6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "ci/ciUtilities.hpp"
27 #include "classfile/javaClasses.hpp"
28 #include "ci/ciObjArray.hpp"
29 #include "asm/register.hpp"
30 #include "compiler/compileLog.hpp"
31 #include "gc/shared/barrierSet.hpp"
32 #include "gc/shared/c2/barrierSetC2.hpp"
33 #include "interpreter/interpreter.hpp"
34 #include "memory/resourceArea.hpp"
35 #include "opto/addnode.hpp"
36 #include "opto/castnode.hpp"
37 #include "opto/convertnode.hpp"
38 #include "opto/graphKit.hpp"
39 #include "opto/idealKit.hpp"
40 #include "opto/intrinsicnode.hpp"
41 #include "opto/locknode.hpp"
42 #include "opto/machnode.hpp"
43 #include "opto/opaquenode.hpp"
44 #include "opto/parse.hpp"
45 #include "opto/rootnode.hpp"
46 #include "opto/runtime.hpp"
47 #include "opto/subtypenode.hpp"
48 #include "runtime/deoptimization.hpp"
49 #include "runtime/sharedRuntime.hpp"
50 #include "utilities/bitMap.inline.hpp"
51 #include "utilities/powerOfTwo.hpp"
52 #include "utilities/growableArray.hpp"
53
54 //----------------------------GraphKit-----------------------------------------
55 // Main utility constructor.
56 GraphKit::GraphKit(JVMState* jvms)
57 : Phase(Phase::Parser),
58 _env(C->env()),
59 _gvn(*C->initial_gvn()),
60 _barrier_set(BarrierSet::barrier_set()->barrier_set_c2())
61 {
62 _exceptions = jvms->map()->next_exception();
63 if (_exceptions != NULL) jvms->map()->set_next_exception(NULL);
64 set_jvms(jvms);
65 }
66
67 // Private constructor for parser.
68 GraphKit::GraphKit()
69 : Phase(Phase::Parser),
70 _env(C->env()),
71 _gvn(*C->initial_gvn()),
72 _barrier_set(BarrierSet::barrier_set()->barrier_set_c2())
73 {
74 _exceptions = NULL;
75 set_map(NULL);
76 debug_only(_sp = -99);
77 debug_only(set_bci(-99));
78 }
79
80
81
82 //---------------------------clean_stack---------------------------------------
83 // Clear away rubbish from the stack area of the JVM state.
84 // This destroys any arguments that may be waiting on the stack.
817 if (PrintMiscellaneous && (Verbose || WizardMode)) {
818 tty->print_cr("Zombie local %d: ", local);
819 jvms->dump();
820 }
821 return false;
822 }
823 }
824 }
825 return true;
826 }
827
828 #endif //ASSERT
829
830 // Helper function for enforcing certain bytecodes to reexecute if deoptimization happens.
831 static bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) {
832 ciMethod* cur_method = jvms->method();
833 int cur_bci = jvms->bci();
834 if (cur_method != NULL && cur_bci != InvocationEntryBci) {
835 Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci);
836 return Interpreter::bytecode_should_reexecute(code) ||
837 (is_anewarray && code == Bytecodes::_multianewarray);
838 // Reexecute _multianewarray bytecode which was replaced with
839 // sequence of [a]newarray. See Parse::do_multianewarray().
840 //
841 // Note: interpreter should not have it set since this optimization
842 // is limited by dimensions and guarded by flag so in some cases
843 // multianewarray() runtime calls will be generated and
844 // the bytecode should not be reexecutes (stack will not be reset).
845 } else {
846 return false;
847 }
848 }
849
850 // Helper function for adding JVMState and debug information to node
851 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) {
852 // Add the safepoint edges to the call (or other safepoint).
853
854 // Make sure dead locals are set to top. This
855 // should help register allocation time and cut down on the size
856 // of the deoptimization information.
857 assert(dead_locals_are_killed(), "garbage in debug info before safepoint");
1077 ciSignature* declared_signature = NULL;
1078 ciMethod* ignored_callee = method()->get_method_at_bci(bci(), ignored_will_link, &declared_signature);
1079 assert(declared_signature != NULL, "cannot be null");
1080 inputs = declared_signature->arg_size_for_bc(code);
1081 int size = declared_signature->return_type()->size();
1082 depth = size - inputs;
1083 }
1084 break;
1085
1086 case Bytecodes::_multianewarray:
1087 {
1088 ciBytecodeStream iter(method());
1089 iter.reset_to_bci(bci());
1090 iter.next();
1091 inputs = iter.get_dimensions();
1092 assert(rsize == 1, "");
1093 depth = rsize - inputs;
1094 }
1095 break;
1096
1097 case Bytecodes::_ireturn:
1098 case Bytecodes::_lreturn:
1099 case Bytecodes::_freturn:
1100 case Bytecodes::_dreturn:
1101 case Bytecodes::_areturn:
1102 assert(rsize == -depth, "");
1103 inputs = rsize;
1104 break;
1105
1106 case Bytecodes::_jsr:
1107 case Bytecodes::_jsr_w:
1108 inputs = 0;
1109 depth = 1; // S.B. depth=1, not zero
1110 break;
1111
1112 default:
1113 // bytecode produces a typed result
1114 inputs = rsize - depth;
1115 assert(inputs >= 0, "");
1116 break;
1159 Node* conv = _gvn.transform( new ConvI2LNode(offset));
1160 Node* mask = _gvn.transform(ConLNode::make((julong) max_juint));
1161 return _gvn.transform( new AndLNode(conv, mask) );
1162 }
1163
1164 Node* GraphKit::ConvL2I(Node* offset) {
1165 // short-circuit a common case
1166 jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot);
1167 if (offset_con != (jlong)Type::OffsetBot) {
1168 return intcon((int) offset_con);
1169 }
1170 return _gvn.transform( new ConvL2INode(offset));
1171 }
1172
1173 //-------------------------load_object_klass-----------------------------------
1174 Node* GraphKit::load_object_klass(Node* obj) {
1175 // Special-case a fresh allocation to avoid building nodes:
1176 Node* akls = AllocateNode::Ideal_klass(obj, &_gvn);
1177 if (akls != NULL) return akls;
1178 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
1179 return _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), k_adr, TypeInstPtr::KLASS));
1180 }
1181
1182 //-------------------------load_array_length-----------------------------------
1183 Node* GraphKit::load_array_length(Node* array) {
1184 // Special-case a fresh allocation to avoid building nodes:
1185 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array, &_gvn);
1186 Node *alen;
1187 if (alloc == NULL) {
1188 Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes());
1189 alen = _gvn.transform( new LoadRangeNode(0, immutable_memory(), r_adr, TypeInt::POS));
1190 } else {
1191 alen = array_ideal_length(alloc, _gvn.type(array)->is_oopptr(), false);
1192 }
1193 return alen;
1194 }
1195
1196 Node* GraphKit::array_ideal_length(AllocateArrayNode* alloc,
1197 const TypeOopPtr* oop_type,
1198 bool replace_length_in_map) {
1199 Node* length = alloc->Ideal_length();
1208 replace_in_map(length, ccast);
1209 }
1210 return ccast;
1211 }
1212 }
1213 return length;
1214 }
1215
1216 //------------------------------do_null_check----------------------------------
1217 // Helper function to do a NULL pointer check. Returned value is
1218 // the incoming address with NULL casted away. You are allowed to use the
1219 // not-null value only if you are control dependent on the test.
1220 #ifndef PRODUCT
1221 extern int explicit_null_checks_inserted,
1222 explicit_null_checks_elided;
1223 #endif
1224 Node* GraphKit::null_check_common(Node* value, BasicType type,
1225 // optional arguments for variations:
1226 bool assert_null,
1227 Node* *null_control,
1228 bool speculative) {
1229 assert(!assert_null || null_control == NULL, "not both at once");
1230 if (stopped()) return top();
1231 NOT_PRODUCT(explicit_null_checks_inserted++);
1232
1233 // Construct NULL check
1234 Node *chk = NULL;
1235 switch(type) {
1236 case T_LONG : chk = new CmpLNode(value, _gvn.zerocon(T_LONG)); break;
1237 case T_INT : chk = new CmpINode(value, _gvn.intcon(0)); break;
1238 case T_ARRAY : // fall through
1239 type = T_OBJECT; // simplify further tests
1240 case T_OBJECT : {
1241 const Type *t = _gvn.type( value );
1242
1243 const TypeOopPtr* tp = t->isa_oopptr();
1244 if (tp != NULL && !tp->is_loaded()
1245 // Only for do_null_check, not any of its siblings:
1246 && !assert_null && null_control == NULL) {
1247 // Usually, any field access or invocation on an unloaded oop type
1248 // will simply fail to link, since the statically linked class is
1249 // likely also to be unloaded. However, in -Xcomp mode, sometimes
1250 // the static class is loaded but the sharper oop type is not.
1251 // Rather than checking for this obscure case in lots of places,
1252 // we simply observe that a null check on an unloaded class
1253 // will always be followed by a nonsense operation, so we
1254 // can just issue the uncommon trap here.
1255 // Our access to the unloaded class will only be correct
1256 // after it has been loaded and initialized, which requires
1257 // a trip through the interpreter.
1316 }
1317 Node *oldcontrol = control();
1318 set_control(cfg);
1319 Node *res = cast_not_null(value);
1320 set_control(oldcontrol);
1321 NOT_PRODUCT(explicit_null_checks_elided++);
1322 return res;
1323 }
1324 cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true);
1325 if (cfg == NULL) break; // Quit at region nodes
1326 depth++;
1327 }
1328 }
1329
1330 //-----------
1331 // Branch to failure if null
1332 float ok_prob = PROB_MAX; // a priori estimate: nulls never happen
1333 Deoptimization::DeoptReason reason;
1334 if (assert_null) {
1335 reason = Deoptimization::reason_null_assert(speculative);
1336 } else if (type == T_OBJECT) {
1337 reason = Deoptimization::reason_null_check(speculative);
1338 } else {
1339 reason = Deoptimization::Reason_div0_check;
1340 }
1341 // %%% Since Reason_unhandled is not recorded on a per-bytecode basis,
1342 // ciMethodData::has_trap_at will return a conservative -1 if any
1343 // must-be-null assertion has failed. This could cause performance
1344 // problems for a method after its first do_null_assert failure.
1345 // Consider using 'Reason_class_check' instead?
1346
1347 // To cause an implicit null check, we set the not-null probability
1348 // to the maximum (PROB_MAX). For an explicit check the probability
1349 // is set to a smaller value.
1350 if (null_control != NULL || too_many_traps(reason)) {
1351 // probability is less likely
1352 ok_prob = PROB_LIKELY_MAG(3);
1353 } else if (!assert_null &&
1354 (ImplicitNullCheckThreshold > 0) &&
1355 method() != NULL &&
1356 (method()->method_data()->trap_count(reason)
1390 }
1391
1392 if (assert_null) {
1393 // Cast obj to null on this path.
1394 replace_in_map(value, zerocon(type));
1395 return zerocon(type);
1396 }
1397
1398 // Cast obj to not-null on this path, if there is no null_control.
1399 // (If there is a null_control, a non-null value may come back to haunt us.)
1400 if (type == T_OBJECT) {
1401 Node* cast = cast_not_null(value, false);
1402 if (null_control == NULL || (*null_control) == top())
1403 replace_in_map(value, cast);
1404 value = cast;
1405 }
1406
1407 return value;
1408 }
1409
1410
1411 //------------------------------cast_not_null----------------------------------
1412 // Cast obj to not-null on this path
1413 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) {
1414 const Type *t = _gvn.type(obj);
1415 const Type *t_not_null = t->join_speculative(TypePtr::NOTNULL);
1416 // Object is already not-null?
1417 if( t == t_not_null ) return obj;
1418
1419 Node *cast = new CastPPNode(obj,t_not_null);
1420 cast->init_req(0, control());
1421 cast = _gvn.transform( cast );
1422
1423 // Scan for instances of 'obj' in the current JVM mapping.
1424 // These instances are known to be not-null after the test.
1425 if (do_replace_in_map)
1426 replace_in_map(obj, cast);
1427
1428 return cast; // Return casted value
1429 }
1430
1431 // Sometimes in intrinsics, we implicitly know an object is not null
1432 // (there's no actual null check) so we can cast it to not null. In
1433 // the course of optimizations, the input to the cast can become null.
1520 // These are layered on top of the factory methods in LoadNode and StoreNode,
1521 // and integrate with the parser's memory state and _gvn engine.
1522 //
1523
1524 // factory methods in "int adr_idx"
1525 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt,
1526 int adr_idx,
1527 MemNode::MemOrd mo,
1528 LoadNode::ControlDependency control_dependency,
1529 bool require_atomic_access,
1530 bool unaligned,
1531 bool mismatched,
1532 bool unsafe,
1533 uint8_t barrier_data) {
1534 assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" );
1535 const TypePtr* adr_type = NULL; // debug-mode-only argument
1536 debug_only(adr_type = C->get_adr_type(adr_idx));
1537 Node* mem = memory(adr_idx);
1538 Node* ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt, mo, control_dependency, require_atomic_access, unaligned, mismatched, unsafe, barrier_data);
1539 ld = _gvn.transform(ld);
1540 if (((bt == T_OBJECT) && C->do_escape_analysis()) || C->eliminate_boxing()) {
1541 // Improve graph before escape analysis and boxing elimination.
1542 record_for_igvn(ld);
1543 }
1544 return ld;
1545 }
1546
1547 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt,
1548 int adr_idx,
1549 MemNode::MemOrd mo,
1550 bool require_atomic_access,
1551 bool unaligned,
1552 bool mismatched,
1553 bool unsafe) {
1554 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
1555 const TypePtr* adr_type = NULL;
1556 debug_only(adr_type = C->get_adr_type(adr_idx));
1557 Node *mem = memory(adr_idx);
1558 Node* st = StoreNode::make(_gvn, ctl, mem, adr, adr_type, val, bt, mo, require_atomic_access);
1559 if (unaligned) {
1560 st->as_Store()->set_unaligned_access();
1564 }
1565 if (unsafe) {
1566 st->as_Store()->set_unsafe_access();
1567 }
1568 st = _gvn.transform(st);
1569 set_memory(st, adr_idx);
1570 // Back-to-back stores can only remove intermediate store with DU info
1571 // so push on worklist for optimizer.
1572 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address))
1573 record_for_igvn(st);
1574
1575 return st;
1576 }
1577
1578 Node* GraphKit::access_store_at(Node* obj,
1579 Node* adr,
1580 const TypePtr* adr_type,
1581 Node* val,
1582 const Type* val_type,
1583 BasicType bt,
1584 DecoratorSet decorators) {
1585 // Transformation of a value which could be NULL pointer (CastPP #NULL)
1586 // could be delayed during Parse (for example, in adjust_map_after_if()).
1587 // Execute transformation here to avoid barrier generation in such case.
1588 if (_gvn.type(val) == TypePtr::NULL_PTR) {
1589 val = _gvn.makecon(TypePtr::NULL_PTR);
1590 }
1591
1592 if (stopped()) {
1593 return top(); // Dead path ?
1594 }
1595
1596 assert(val != NULL, "not dead path");
1597
1598 C2AccessValuePtr addr(adr, adr_type);
1599 C2AccessValue value(val, val_type);
1600 C2ParseAccess access(this, decorators | C2_WRITE_ACCESS, bt, obj, addr);
1601 if (access.is_raw()) {
1602 return _barrier_set->BarrierSetC2::store_at(access, value);
1603 } else {
1604 return _barrier_set->store_at(access, value);
1605 }
1606 }
1607
1608 Node* GraphKit::access_load_at(Node* obj, // containing obj
1609 Node* adr, // actual address to store val at
1610 const TypePtr* adr_type,
1611 const Type* val_type,
1612 BasicType bt,
1613 DecoratorSet decorators) {
1614 if (stopped()) {
1615 return top(); // Dead path ?
1616 }
1617
1618 C2AccessValuePtr addr(adr, adr_type);
1619 C2ParseAccess access(this, decorators | C2_READ_ACCESS, bt, obj, addr);
1620 if (access.is_raw()) {
1621 return _barrier_set->BarrierSetC2::load_at(access, val_type);
1622 } else {
1623 return _barrier_set->load_at(access, val_type);
1624 }
1625 }
1626
1627 Node* GraphKit::access_load(Node* adr, // actual address to load val at
1628 const Type* val_type,
1629 BasicType bt,
1630 DecoratorSet decorators) {
1631 if (stopped()) {
1632 return top(); // Dead path ?
1633 }
1634
1635 C2AccessValuePtr addr(adr, adr->bottom_type()->is_ptr());
1636 C2ParseAccess access(this, decorators | C2_READ_ACCESS, bt, NULL, addr);
1637 if (access.is_raw()) {
1638 return _barrier_set->BarrierSetC2::load_at(access, val_type);
1639 } else {
1704 Node* new_val,
1705 const Type* value_type,
1706 BasicType bt,
1707 DecoratorSet decorators) {
1708 C2AccessValuePtr addr(adr, adr_type);
1709 C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS, bt, obj, addr, alias_idx);
1710 if (access.is_raw()) {
1711 return _barrier_set->BarrierSetC2::atomic_add_at(access, new_val, value_type);
1712 } else {
1713 return _barrier_set->atomic_add_at(access, new_val, value_type);
1714 }
1715 }
1716
1717 void GraphKit::access_clone(Node* src, Node* dst, Node* size, bool is_array) {
1718 return _barrier_set->clone(this, src, dst, size, is_array);
1719 }
1720
1721 //-------------------------array_element_address-------------------------
1722 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt,
1723 const TypeInt* sizetype, Node* ctrl) {
1724 uint shift = exact_log2(type2aelembytes(elembt));
1725 uint header = arrayOopDesc::base_offset_in_bytes(elembt);
1726
1727 // short-circuit a common case (saves lots of confusing waste motion)
1728 jint idx_con = find_int_con(idx, -1);
1729 if (idx_con >= 0) {
1730 intptr_t offset = header + ((intptr_t)idx_con << shift);
1731 return basic_plus_adr(ary, offset);
1732 }
1733
1734 // must be correct type for alignment purposes
1735 Node* base = basic_plus_adr(ary, header);
1736 idx = Compile::conv_I2X_index(&_gvn, idx, sizetype, ctrl);
1737 Node* scale = _gvn.transform( new LShiftXNode(idx, intcon(shift)) );
1738 return basic_plus_adr(ary, base, scale);
1739 }
1740
1741 //-------------------------load_array_element-------------------------
1742 Node* GraphKit::load_array_element(Node* ary, Node* idx, const TypeAryPtr* arytype, bool set_ctrl) {
1743 const Type* elemtype = arytype->elem();
1744 BasicType elembt = elemtype->array_element_basic_type();
1745 Node* adr = array_element_address(ary, idx, elembt, arytype->size());
1746 if (elembt == T_NARROWOOP) {
1747 elembt = T_OBJECT; // To satisfy switch in LoadNode::make()
1748 }
1749 Node* ld = access_load_at(ary, adr, arytype, elemtype, elembt,
1750 IN_HEAP | IS_ARRAY | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0));
1751 return ld;
1752 }
1753
1754 //-------------------------set_arguments_for_java_call-------------------------
1755 // Arguments (pre-popped from the stack) are taken from the JVMS.
1756 void GraphKit::set_arguments_for_java_call(CallJavaNode* call) {
1757 // Add the call arguments:
1758 uint nargs = call->method()->arg_size();
1759 for (uint i = 0; i < nargs; i++) {
1760 Node* arg = argument(i);
1761 call->init_req(i + TypeFunc::Parms, arg);
1762 }
1763 }
1764
1765 //---------------------------set_edges_for_java_call---------------------------
1766 // Connect a newly created call into the current JVMS.
1767 // A return value node (if any) is returned from set_edges_for_java_call.
1768 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) {
1769
1770 // Add the predefined inputs:
1771 call->init_req( TypeFunc::Control, control() );
1772 call->init_req( TypeFunc::I_O , i_o() );
1773 call->init_req( TypeFunc::Memory , reset_memory() );
1774 call->init_req( TypeFunc::FramePtr, frameptr() );
1775 call->init_req( TypeFunc::ReturnAdr, top() );
1776
1777 add_safepoint_edges(call, must_throw);
1778
1779 Node* xcall = _gvn.transform(call);
1780
1781 if (xcall == top()) {
1782 set_control(top());
1783 return;
1784 }
1785 assert(xcall == call, "call identity is stable");
1786
1787 // Re-use the current map to produce the result.
1788
1789 set_control(_gvn.transform(new ProjNode(call, TypeFunc::Control)));
1790 set_i_o( _gvn.transform(new ProjNode(call, TypeFunc::I_O , separate_io_proj)));
1791 set_all_memory_call(xcall, separate_io_proj);
1792
1793 //return xcall; // no need, caller already has it
1794 }
1795
1796 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj, bool deoptimize) {
1797 if (stopped()) return top(); // maybe the call folded up?
1798
1799 // Capture the return value, if any.
1800 Node* ret;
1801 if (call->method() == NULL ||
1802 call->method()->return_type()->basic_type() == T_VOID)
1803 ret = top();
1804 else ret = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
1805
1806 // Note: Since any out-of-line call can produce an exception,
1807 // we always insert an I_O projection from the call into the result.
1808
1809 make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj, deoptimize);
1810
1811 if (separate_io_proj) {
1812 // The caller requested separate projections be used by the fall
1813 // through and exceptional paths, so replace the projections for
1814 // the fall through path.
1815 set_i_o(_gvn.transform( new ProjNode(call, TypeFunc::I_O) ));
1816 set_all_memory(_gvn.transform( new ProjNode(call, TypeFunc::Memory) ));
1817 }
1818 return ret;
1819 }
1820
1821 //--------------------set_predefined_input_for_runtime_call--------------------
1822 // Reading and setting the memory state is way conservative here.
1823 // The real problem is that I am not doing real Type analysis on memory,
1824 // so I cannot distinguish card mark stores from other stores. Across a GC
1825 // point the Store Barrier and the card mark memory has to agree. I cannot
1826 // have a card mark store and its barrier split across the GC point from
1827 // either above or below. Here I get that to happen by reading ALL of memory.
1828 // A better answer would be to separate out card marks from other memory.
1829 // For now, return the input memory state, so that it can be reused
1830 // after the call, if this call has restricted memory effects.
1831 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call, Node* narrow_mem) {
1832 // Set fixed predefined input arguments
1833 Node* memory = reset_memory();
1834 Node* m = narrow_mem == NULL ? memory : narrow_mem;
1835 call->init_req( TypeFunc::Control, control() );
1836 call->init_req( TypeFunc::I_O, top() ); // does no i/o
1837 call->init_req( TypeFunc::Memory, m ); // may gc ptrs
1888 if (use->is_MergeMem()) {
1889 wl.push(use);
1890 }
1891 }
1892 }
1893
1894 // Replace the call with the current state of the kit.
1895 void GraphKit::replace_call(CallNode* call, Node* result, bool do_replaced_nodes) {
1896 JVMState* ejvms = NULL;
1897 if (has_exceptions()) {
1898 ejvms = transfer_exceptions_into_jvms();
1899 }
1900
1901 ReplacedNodes replaced_nodes = map()->replaced_nodes();
1902 ReplacedNodes replaced_nodes_exception;
1903 Node* ex_ctl = top();
1904
1905 SafePointNode* final_state = stop();
1906
1907 // Find all the needed outputs of this call
1908 CallProjections callprojs;
1909 call->extract_projections(&callprojs, true);
1910
1911 Unique_Node_List wl;
1912 Node* init_mem = call->in(TypeFunc::Memory);
1913 Node* final_mem = final_state->in(TypeFunc::Memory);
1914 Node* final_ctl = final_state->in(TypeFunc::Control);
1915 Node* final_io = final_state->in(TypeFunc::I_O);
1916
1917 // Replace all the old call edges with the edges from the inlining result
1918 if (callprojs.fallthrough_catchproj != NULL) {
1919 C->gvn_replace_by(callprojs.fallthrough_catchproj, final_ctl);
1920 }
1921 if (callprojs.fallthrough_memproj != NULL) {
1922 if (final_mem->is_MergeMem()) {
1923 // Parser's exits MergeMem was not transformed but may be optimized
1924 final_mem = _gvn.transform(final_mem);
1925 }
1926 C->gvn_replace_by(callprojs.fallthrough_memproj, final_mem);
1927 add_mergemem_users_to_worklist(wl, final_mem);
1928 }
1929 if (callprojs.fallthrough_ioproj != NULL) {
1930 C->gvn_replace_by(callprojs.fallthrough_ioproj, final_io);
1931 }
1932
1933 // Replace the result with the new result if it exists and is used
1934 if (callprojs.resproj != NULL && result != NULL) {
1935 C->gvn_replace_by(callprojs.resproj, result);
1936 }
1937
1938 if (ejvms == NULL) {
1939 // No exception edges to simply kill off those paths
1940 if (callprojs.catchall_catchproj != NULL) {
1941 C->gvn_replace_by(callprojs.catchall_catchproj, C->top());
1942 }
1943 if (callprojs.catchall_memproj != NULL) {
1944 C->gvn_replace_by(callprojs.catchall_memproj, C->top());
1945 }
1946 if (callprojs.catchall_ioproj != NULL) {
1947 C->gvn_replace_by(callprojs.catchall_ioproj, C->top());
1948 }
1949 // Replace the old exception object with top
1950 if (callprojs.exobj != NULL) {
1951 C->gvn_replace_by(callprojs.exobj, C->top());
1952 }
1953 } else {
1954 GraphKit ekit(ejvms);
1955
1956 // Load my combined exception state into the kit, with all phis transformed:
1957 SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states();
1958 replaced_nodes_exception = ex_map->replaced_nodes();
1959
1960 Node* ex_oop = ekit.use_exception_state(ex_map);
1961
1962 if (callprojs.catchall_catchproj != NULL) {
1963 C->gvn_replace_by(callprojs.catchall_catchproj, ekit.control());
1964 ex_ctl = ekit.control();
1965 }
1966 if (callprojs.catchall_memproj != NULL) {
1967 Node* ex_mem = ekit.reset_memory();
1968 C->gvn_replace_by(callprojs.catchall_memproj, ex_mem);
1969 add_mergemem_users_to_worklist(wl, ex_mem);
1970 }
1971 if (callprojs.catchall_ioproj != NULL) {
1972 C->gvn_replace_by(callprojs.catchall_ioproj, ekit.i_o());
1973 }
1974
1975 // Replace the old exception object with the newly created one
1976 if (callprojs.exobj != NULL) {
1977 C->gvn_replace_by(callprojs.exobj, ex_oop);
1978 }
1979 }
1980
1981 // Disconnect the call from the graph
1982 call->disconnect_inputs(C);
1983 C->gvn_replace_by(call, C->top());
1984
1985 // Clean up any MergeMems that feed other MergeMems since the
1986 // optimizer doesn't like that.
1987 while (wl.size() > 0) {
1988 _gvn.transform(wl.pop());
1989 }
1990
1991 if (callprojs.fallthrough_catchproj != NULL && !final_ctl->is_top() && do_replaced_nodes) {
1992 replaced_nodes.apply(C, final_ctl);
1993 }
1994 if (!ex_ctl->is_top() && do_replaced_nodes) {
1995 replaced_nodes_exception.apply(C, ex_ctl);
1996 }
1997 }
1998
1999
2000 //------------------------------increment_counter------------------------------
2001 // for statistics: increment a VM counter by 1
2002
2003 void GraphKit::increment_counter(address counter_addr) {
2004 Node* adr1 = makecon(TypeRawPtr::make(counter_addr));
2005 increment_counter(adr1);
2006 }
2007
2008 void GraphKit::increment_counter(Node* counter_addr) {
2009 int adr_type = Compile::AliasIdxRaw;
2010 Node* ctrl = control();
2011 Node* cnt = make_load(ctrl, counter_addr, TypeLong::LONG, T_LONG, adr_type, MemNode::unordered);
2170 *
2171 * @param n node that the type applies to
2172 * @param exact_kls type from profiling
2173 * @param maybe_null did profiling see null?
2174 *
2175 * @return node with improved type
2176 */
2177 Node* GraphKit::record_profile_for_speculation(Node* n, ciKlass* exact_kls, ProfilePtrKind ptr_kind) {
2178 const Type* current_type = _gvn.type(n);
2179 assert(UseTypeSpeculation, "type speculation must be on");
2180
2181 const TypePtr* speculative = current_type->speculative();
2182
2183 // Should the klass from the profile be recorded in the speculative type?
2184 if (current_type->would_improve_type(exact_kls, jvms()->depth())) {
2185 const TypeKlassPtr* tklass = TypeKlassPtr::make(exact_kls);
2186 const TypeOopPtr* xtype = tklass->as_instance_type();
2187 assert(xtype->klass_is_exact(), "Should be exact");
2188 // Any reason to believe n is not null (from this profiling or a previous one)?
2189 assert(ptr_kind != ProfileAlwaysNull, "impossible here");
2190 const TypePtr* ptr = (ptr_kind == ProfileMaybeNull && current_type->speculative_maybe_null()) ? TypePtr::BOTTOM : TypePtr::NOTNULL;
2191 // record the new speculative type's depth
2192 speculative = xtype->cast_to_ptr_type(ptr->ptr())->is_ptr();
2193 speculative = speculative->with_inline_depth(jvms()->depth());
2194 } else if (current_type->would_improve_ptr(ptr_kind)) {
2195 // Profiling report that null was never seen so we can change the
2196 // speculative type to non null ptr.
2197 if (ptr_kind == ProfileAlwaysNull) {
2198 speculative = TypePtr::NULL_PTR;
2199 } else {
2200 assert(ptr_kind == ProfileNeverNull, "nothing else is an improvement");
2201 const TypePtr* ptr = TypePtr::NOTNULL;
2202 if (speculative != NULL) {
2203 speculative = speculative->cast_to_ptr_type(ptr->ptr())->is_ptr();
2204 } else {
2205 speculative = ptr;
2206 }
2207 }
2208 }
2209
2210 if (speculative != current_type->speculative()) {
2211 // Build a type with a speculative type (what we think we know
2212 // about the type but will need a guard when we use it)
2213 const TypeOopPtr* spec_type = TypeOopPtr::make(TypePtr::BotPTR, Type::OffsetBot, TypeOopPtr::InstanceBot, speculative);
2214 // We're changing the type, we need a new CheckCast node to carry
2215 // the new type. The new type depends on the control: what
2216 // profiling tells us is only valid from here as far as we can
2217 // tell.
2218 Node* cast = new CheckCastPPNode(control(), n, current_type->remove_speculative()->join_speculative(spec_type));
2219 cast = _gvn.transform(cast);
2220 replace_in_map(n, cast);
2221 n = cast;
2222 }
2223
2224 return n;
2225 }
2226
2227 /**
2228 * Record profiling data from receiver profiling at an invoke with the
2229 * type system so that it can propagate it (speculation)
2230 *
2231 * @param n receiver node
2232 *
2233 * @return node with improved type
2234 */
2235 Node* GraphKit::record_profiled_receiver_for_speculation(Node* n) {
2236 if (!UseTypeSpeculation) {
2237 return n;
2238 }
2239 ciKlass* exact_kls = profile_has_unique_klass();
2240 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2241 if ((java_bc() == Bytecodes::_checkcast ||
2242 java_bc() == Bytecodes::_instanceof ||
2243 java_bc() == Bytecodes::_aastore) &&
2244 method()->method_data()->is_mature()) {
2245 ciProfileData* data = method()->method_data()->bci_to_data(bci());
2246 if (data != NULL) {
2247 if (!data->as_BitData()->null_seen()) {
2248 ptr_kind = ProfileNeverNull;
2249 } else {
2250 assert(data->is_ReceiverTypeData(), "bad profile data type");
2251 ciReceiverTypeData* call = (ciReceiverTypeData*)data->as_ReceiverTypeData();
2252 uint i = 0;
2253 for (; i < call->row_limit(); i++) {
2254 ciKlass* receiver = call->receiver(i);
2255 if (receiver != NULL) {
2256 break;
2257 }
2258 }
2259 ptr_kind = (i == call->row_limit()) ? ProfileAlwaysNull : ProfileMaybeNull;
2260 }
2261 }
2262 }
2263 return record_profile_for_speculation(n, exact_kls, ptr_kind);
2264 }
2265
2266 /**
2267 * Record profiling data from argument profiling at an invoke with the
2268 * type system so that it can propagate it (speculation)
2269 *
2270 * @param dest_method target method for the call
2271 * @param bc what invoke bytecode is this?
2272 */
2273 void GraphKit::record_profiled_arguments_for_speculation(ciMethod* dest_method, Bytecodes::Code bc) {
2274 if (!UseTypeSpeculation) {
2275 return;
2276 }
2277 const TypeFunc* tf = TypeFunc::make(dest_method);
2278 int nargs = tf->domain()->cnt() - TypeFunc::Parms;
2279 int skip = Bytecodes::has_receiver(bc) ? 1 : 0;
2280 for (int j = skip, i = 0; j < nargs && i < TypeProfileArgsLimit; j++) {
2281 const Type *targ = tf->domain()->field_at(j + TypeFunc::Parms);
2282 if (is_reference_type(targ->basic_type())) {
2283 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2284 ciKlass* better_type = NULL;
2285 if (method()->argument_profiled_type(bci(), i, better_type, ptr_kind)) {
2286 record_profile_for_speculation(argument(j), better_type, ptr_kind);
2287 }
2288 i++;
2289 }
2290 }
2291 }
2292
2293 /**
2294 * Record profiling data from parameter profiling at an invoke with
2295 * the type system so that it can propagate it (speculation)
2296 */
2297 void GraphKit::record_profiled_parameters_for_speculation() {
2298 if (!UseTypeSpeculation) {
2299 return;
2300 }
2301 for (int i = 0, j = 0; i < method()->arg_size() ; i++) {
2315 * the type system so that it can propagate it (speculation)
2316 */
2317 void GraphKit::record_profiled_return_for_speculation() {
2318 if (!UseTypeSpeculation) {
2319 return;
2320 }
2321 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2322 ciKlass* better_type = NULL;
2323 if (method()->return_profiled_type(bci(), better_type, ptr_kind)) {
2324 // If profiling reports a single type for the return value,
2325 // feed it to the type system so it can propagate it as a
2326 // speculative type
2327 record_profile_for_speculation(stack(sp()-1), better_type, ptr_kind);
2328 }
2329 }
2330
2331 void GraphKit::round_double_arguments(ciMethod* dest_method) {
2332 if (Matcher::strict_fp_requires_explicit_rounding) {
2333 // (Note: TypeFunc::make has a cache that makes this fast.)
2334 const TypeFunc* tf = TypeFunc::make(dest_method);
2335 int nargs = tf->domain()->cnt() - TypeFunc::Parms;
2336 for (int j = 0; j < nargs; j++) {
2337 const Type *targ = tf->domain()->field_at(j + TypeFunc::Parms);
2338 if (targ->basic_type() == T_DOUBLE) {
2339 // If any parameters are doubles, they must be rounded before
2340 // the call, dprecision_rounding does gvn.transform
2341 Node *arg = argument(j);
2342 arg = dprecision_rounding(arg);
2343 set_argument(j, arg);
2344 }
2345 }
2346 }
2347 }
2348
2349 // rounding for strict float precision conformance
2350 Node* GraphKit::precision_rounding(Node* n) {
2351 if (Matcher::strict_fp_requires_explicit_rounding) {
2352 #ifdef IA32
2353 if (UseSSE == 0) {
2354 return _gvn.transform(new RoundFloatNode(0, n));
2355 }
2356 #else
2357 Unimplemented();
2466 // The first NULL ends the list.
2467 Node* parm0, Node* parm1,
2468 Node* parm2, Node* parm3,
2469 Node* parm4, Node* parm5,
2470 Node* parm6, Node* parm7) {
2471 assert(call_addr != NULL, "must not call NULL targets");
2472
2473 // Slow-path call
2474 bool is_leaf = !(flags & RC_NO_LEAF);
2475 bool has_io = (!is_leaf && !(flags & RC_NO_IO));
2476 if (call_name == NULL) {
2477 assert(!is_leaf, "must supply name for leaf");
2478 call_name = OptoRuntime::stub_name(call_addr);
2479 }
2480 CallNode* call;
2481 if (!is_leaf) {
2482 call = new CallStaticJavaNode(call_type, call_addr, call_name, adr_type);
2483 } else if (flags & RC_NO_FP) {
2484 call = new CallLeafNoFPNode(call_type, call_addr, call_name, adr_type);
2485 } else if (flags & RC_VECTOR){
2486 uint num_bits = call_type->range()->field_at(TypeFunc::Parms)->is_vect()->length_in_bytes() * BitsPerByte;
2487 call = new CallLeafVectorNode(call_type, call_addr, call_name, adr_type, num_bits);
2488 } else {
2489 call = new CallLeafNode(call_type, call_addr, call_name, adr_type);
2490 }
2491
2492 // The following is similar to set_edges_for_java_call,
2493 // except that the memory effects of the call are restricted to AliasIdxRaw.
2494
2495 // Slow path call has no side-effects, uses few values
2496 bool wide_in = !(flags & RC_NARROW_MEM);
2497 bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot);
2498
2499 Node* prev_mem = NULL;
2500 if (wide_in) {
2501 prev_mem = set_predefined_input_for_runtime_call(call);
2502 } else {
2503 assert(!wide_out, "narrow in => narrow out");
2504 Node* narrow_mem = memory(adr_type);
2505 prev_mem = set_predefined_input_for_runtime_call(call, narrow_mem);
2506 }
2546
2547 if (has_io) {
2548 set_i_o(_gvn.transform(new ProjNode(call, TypeFunc::I_O)));
2549 }
2550 return call;
2551
2552 }
2553
2554 // i2b
2555 Node* GraphKit::sign_extend_byte(Node* in) {
2556 Node* tmp = _gvn.transform(new LShiftINode(in, _gvn.intcon(24)));
2557 return _gvn.transform(new RShiftINode(tmp, _gvn.intcon(24)));
2558 }
2559
2560 // i2s
2561 Node* GraphKit::sign_extend_short(Node* in) {
2562 Node* tmp = _gvn.transform(new LShiftINode(in, _gvn.intcon(16)));
2563 return _gvn.transform(new RShiftINode(tmp, _gvn.intcon(16)));
2564 }
2565
2566 //------------------------------merge_memory-----------------------------------
2567 // Merge memory from one path into the current memory state.
2568 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) {
2569 for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) {
2570 Node* old_slice = mms.force_memory();
2571 Node* new_slice = mms.memory2();
2572 if (old_slice != new_slice) {
2573 PhiNode* phi;
2574 if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region) {
2575 if (mms.is_empty()) {
2576 // clone base memory Phi's inputs for this memory slice
2577 assert(old_slice == mms.base_memory(), "sanity");
2578 phi = PhiNode::make(region, NULL, Type::MEMORY, mms.adr_type(C));
2579 _gvn.set_type(phi, Type::MEMORY);
2580 for (uint i = 1; i < phi->req(); i++) {
2581 phi->init_req(i, old_slice->in(i));
2582 }
2583 } else {
2584 phi = old_slice->as_Phi(); // Phi was generated already
2585 }
2797
2798 // Now do a linear scan of the secondary super-klass array. Again, no real
2799 // performance impact (too rare) but it's gotta be done.
2800 // Since the code is rarely used, there is no penalty for moving it
2801 // out of line, and it can only improve I-cache density.
2802 // The decision to inline or out-of-line this final check is platform
2803 // dependent, and is found in the AD file definition of PartialSubtypeCheck.
2804 Node* psc = gvn.transform(
2805 new PartialSubtypeCheckNode(*ctrl, subklass, superklass));
2806
2807 IfNode *iff4 = gen_subtype_check_compare(*ctrl, psc, gvn.zerocon(T_OBJECT), BoolTest::ne, PROB_FAIR, gvn, T_ADDRESS);
2808 r_not_subtype->init_req(2, gvn.transform(new IfTrueNode (iff4)));
2809 r_ok_subtype ->init_req(3, gvn.transform(new IfFalseNode(iff4)));
2810
2811 // Return false path; set default control to true path.
2812 *ctrl = gvn.transform(r_ok_subtype);
2813 return gvn.transform(r_not_subtype);
2814 }
2815
2816 Node* GraphKit::gen_subtype_check(Node* obj_or_subklass, Node* superklass) {
2817 bool expand_subtype_check = C->post_loop_opts_phase() || // macro node expansion is over
2818 ExpandSubTypeCheckAtParseTime; // forced expansion
2819 if (expand_subtype_check) {
2820 MergeMemNode* mem = merged_memory();
2821 Node* ctrl = control();
2822 Node* subklass = obj_or_subklass;
2823 if (!_gvn.type(obj_or_subklass)->isa_klassptr()) {
2824 subklass = load_object_klass(obj_or_subklass);
2825 }
2826
2827 Node* n = Phase::gen_subtype_check(subklass, superklass, &ctrl, mem, _gvn);
2828 set_control(ctrl);
2829 return n;
2830 }
2831
2832 Node* check = _gvn.transform(new SubTypeCheckNode(C, obj_or_subklass, superklass));
2833 Node* bol = _gvn.transform(new BoolNode(check, BoolTest::eq));
2834 IfNode* iff = create_and_xform_if(control(), bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
2835 set_control(_gvn.transform(new IfTrueNode(iff)));
2836 return _gvn.transform(new IfFalseNode(iff));
2837 }
2838
2839 // Profile-driven exact type check:
2840 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass,
2841 float prob,
2842 Node* *casted_receiver) {
2843 assert(!klass->is_interface(), "no exact type check on interfaces");
2844
2845 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass);
2846 Node* recv_klass = load_object_klass(receiver);
2847 Node* want_klass = makecon(tklass);
2848 Node* cmp = _gvn.transform(new CmpPNode(recv_klass, want_klass));
2849 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
2850 IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN);
2851 set_control( _gvn.transform(new IfTrueNode (iff)));
2852 Node* fail = _gvn.transform(new IfFalseNode(iff));
2853
2854 if (!stopped()) {
2855 const TypeOopPtr* receiver_type = _gvn.type(receiver)->isa_oopptr();
2856 const TypeOopPtr* recvx_type = tklass->as_instance_type();
2857 assert(recvx_type->klass_is_exact(), "");
2858
2859 if (!receiver_type->higher_equal(recvx_type)) { // ignore redundant casts
2860 // Subsume downstream occurrences of receiver with a cast to
2861 // recv_xtype, since now we know what the type will be.
2862 Node* cast = new CheckCastPPNode(control(), receiver, recvx_type);
2863 (*casted_receiver) = _gvn.transform(cast);
2864 // (User must make the replace_in_map call.)
2865 }
2866 }
2867
2868 return fail;
2869 }
2870
2871 //------------------------------subtype_check_receiver-------------------------
2872 Node* GraphKit::subtype_check_receiver(Node* receiver, ciKlass* klass,
2873 Node** casted_receiver) {
2874 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass);
2875 Node* want_klass = makecon(tklass);
2876
2877 Node* slow_ctl = gen_subtype_check(receiver, want_klass);
2878
2879 // Ignore interface type information until interface types are properly tracked.
2880 if (!stopped() && !klass->is_interface()) {
2881 const TypeOopPtr* receiver_type = _gvn.type(receiver)->isa_oopptr();
2882 const TypeOopPtr* recv_type = tklass->cast_to_exactness(false)->is_klassptr()->as_instance_type();
2883 if (!receiver_type->higher_equal(recv_type)) { // ignore redundant casts
2884 Node* cast = new CheckCastPPNode(control(), receiver, recv_type);
2885 (*casted_receiver) = _gvn.transform(cast);
2886 }
2887 }
2888
2889 return slow_ctl;
2890 }
2891
2892 //------------------------------seems_never_null-------------------------------
2893 // Use null_seen information if it is available from the profile.
2894 // If we see an unexpected null at a type check we record it and force a
2895 // recompile; the offending check will be recompiled to handle NULLs.
2896 // If we see several offending BCIs, then all checks in the
2897 // method will be recompiled.
2898 bool GraphKit::seems_never_null(Node* obj, ciProfileData* data, bool& speculating) {
2899 speculating = !_gvn.type(obj)->speculative_maybe_null();
2900 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculating);
2901 if (UncommonNullCast // Cutout for this technique
2902 && obj != null() // And not the -Xcomp stupid case?
2903 && !too_many_traps(reason)
2904 ) {
2905 if (speculating) {
2906 return true;
2907 }
2908 if (data == NULL)
2909 // Edge case: no mature data. Be optimistic here.
2910 return true;
2911 // If the profile has not seen a null, assume it won't happen.
2912 assert(java_bc() == Bytecodes::_checkcast ||
2913 java_bc() == Bytecodes::_instanceof ||
2914 java_bc() == Bytecodes::_aastore, "MDO must collect null_seen bit here");
2915 return !data->as_BitData()->null_seen();
2916 }
2917 speculating = false;
2918 return false;
2919 }
2920
2921 void GraphKit::guard_klass_being_initialized(Node* klass) {
2922 int init_state_off = in_bytes(InstanceKlass::init_state_offset());
2923 Node* adr = basic_plus_adr(top(), klass, init_state_off);
2924 Node* init_state = LoadNode::make(_gvn, NULL, immutable_memory(), adr,
2925 adr->bottom_type()->is_ptr(), TypeInt::BYTE,
2926 T_BYTE, MemNode::unordered);
2927 init_state = _gvn.transform(init_state);
2928
2929 Node* being_initialized_state = makecon(TypeInt::make(InstanceKlass::being_initialized));
2930
2931 Node* chk = _gvn.transform(new CmpINode(being_initialized_state, init_state));
2932 Node* tst = _gvn.transform(new BoolNode(chk, BoolTest::eq));
2933
2934 { BuildCutout unless(this, tst, PROB_MAX);
2974
2975 //------------------------maybe_cast_profiled_receiver-------------------------
2976 // If the profile has seen exactly one type, narrow to exactly that type.
2977 // Subsequent type checks will always fold up.
2978 Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj,
2979 const TypeKlassPtr* require_klass,
2980 ciKlass* spec_klass,
2981 bool safe_for_replace) {
2982 if (!UseTypeProfile || !TypeProfileCasts) return NULL;
2983
2984 Deoptimization::DeoptReason reason = Deoptimization::reason_class_check(spec_klass != NULL);
2985
2986 // Make sure we haven't already deoptimized from this tactic.
2987 if (too_many_traps_or_recompiles(reason))
2988 return NULL;
2989
2990 // (No, this isn't a call, but it's enough like a virtual call
2991 // to use the same ciMethod accessor to get the profile info...)
2992 // If we have a speculative type use it instead of profiling (which
2993 // may not help us)
2994 ciKlass* exact_kls = spec_klass == NULL ? profile_has_unique_klass() : spec_klass;
2995 if (exact_kls != NULL) {// no cast failures here
2996 if (require_klass == NULL ||
2997 C->static_subtype_check(require_klass, TypeKlassPtr::make(exact_kls)) == Compile::SSC_always_true) {
2998 // If we narrow the type to match what the type profile sees or
2999 // the speculative type, we can then remove the rest of the
3000 // cast.
3001 // This is a win, even if the exact_kls is very specific,
3002 // because downstream operations, such as method calls,
3003 // will often benefit from the sharper type.
3004 Node* exact_obj = not_null_obj; // will get updated in place...
3005 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0,
3006 &exact_obj);
3007 { PreserveJVMState pjvms(this);
3008 set_control(slow_ctl);
3009 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile);
3010 }
3011 if (safe_for_replace) {
3012 replace_in_map(not_null_obj, exact_obj);
3013 }
3014 return exact_obj;
3079 // and the reflective instance-of call.
3080 Node* GraphKit::gen_instanceof(Node* obj, Node* superklass, bool safe_for_replace) {
3081 kill_dead_locals(); // Benefit all the uncommon traps
3082 assert( !stopped(), "dead parse path should be checked in callers" );
3083 assert(!TypePtr::NULL_PTR->higher_equal(_gvn.type(superklass)->is_klassptr()),
3084 "must check for not-null not-dead klass in callers");
3085
3086 // Make the merge point
3087 enum { _obj_path = 1, _fail_path, _null_path, PATH_LIMIT };
3088 RegionNode* region = new RegionNode(PATH_LIMIT);
3089 Node* phi = new PhiNode(region, TypeInt::BOOL);
3090 C->set_has_split_ifs(true); // Has chance for split-if optimization
3091
3092 ciProfileData* data = NULL;
3093 if (java_bc() == Bytecodes::_instanceof) { // Only for the bytecode
3094 data = method()->method_data()->bci_to_data(bci());
3095 }
3096 bool speculative_not_null = false;
3097 bool never_see_null = (ProfileDynamicTypes // aggressive use of profile
3098 && seems_never_null(obj, data, speculative_not_null));
3099
3100 // Null check; get casted pointer; set region slot 3
3101 Node* null_ctl = top();
3102 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null);
3103
3104 // If not_null_obj is dead, only null-path is taken
3105 if (stopped()) { // Doing instance-of on a NULL?
3106 set_control(null_ctl);
3107 return intcon(0);
3108 }
3109 region->init_req(_null_path, null_ctl);
3110 phi ->init_req(_null_path, intcon(0)); // Set null path value
3111 if (null_ctl == top()) {
3112 // Do this eagerly, so that pattern matches like is_diamond_phi
3113 // will work even during parsing.
3114 assert(_null_path == PATH_LIMIT-1, "delete last");
3115 region->del_req(_null_path);
3116 phi ->del_req(_null_path);
3117 }
3118
3119 // Do we know the type check always succeed?
3120 bool known_statically = false;
3121 if (_gvn.type(superklass)->singleton()) {
3122 const TypeKlassPtr* superk = _gvn.type(superklass)->is_klassptr();
3123 const TypeKlassPtr* subk = _gvn.type(obj)->is_oopptr()->as_klass_type();
3124 if (subk->is_loaded()) {
3125 int static_res = C->static_subtype_check(superk, subk);
3126 known_statically = (static_res == Compile::SSC_always_true || static_res == Compile::SSC_always_false);
3127 }
3128 }
3129
3130 if (!known_statically) {
3131 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3132 // We may not have profiling here or it may not help us. If we
3133 // have a speculative type use it to perform an exact cast.
3134 ciKlass* spec_obj_type = obj_type->speculative_type();
3135 if (spec_obj_type != NULL || (ProfileDynamicTypes && data != NULL)) {
3136 Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, NULL, spec_obj_type, safe_for_replace);
3137 if (stopped()) { // Profile disagrees with this path.
3138 set_control(null_ctl); // Null is the only remaining possibility.
3139 return intcon(0);
3140 }
3141 if (cast_obj != NULL) {
3142 not_null_obj = cast_obj;
3143 }
3144 }
3145 }
3146
3147 // Generate the subtype check
3148 Node* not_subtype_ctrl = gen_subtype_check(not_null_obj, superklass);
3149
3150 // Plug in the success path to the general merge in slot 1.
3151 region->init_req(_obj_path, control());
3152 phi ->init_req(_obj_path, intcon(1));
3153
3154 // Plug in the failing path to the general merge in slot 2.
3155 region->init_req(_fail_path, not_subtype_ctrl);
3156 phi ->init_req(_fail_path, intcon(0));
3157
3158 // Return final merged results
3159 set_control( _gvn.transform(region) );
3160 record_for_igvn(region);
3161
3162 // If we know the type check always succeeds then we don't use the
3163 // profiling data at this bytecode. Don't lose it, feed it to the
3164 // type system as a speculative type.
3165 if (safe_for_replace) {
3166 Node* casted_obj = record_profiled_receiver_for_speculation(obj);
3167 replace_in_map(obj, casted_obj);
3168 }
3169
3170 return _gvn.transform(phi);
3171 }
3172
3173 //-------------------------------gen_checkcast---------------------------------
3174 // Generate a checkcast idiom. Used by both the checkcast bytecode and the
3175 // array store bytecode. Stack must be as-if BEFORE doing the bytecode so the
3176 // uncommon-trap paths work. Adjust stack after this call.
3177 // If failure_control is supplied and not null, it is filled in with
3178 // the control edge for the cast failure. Otherwise, an appropriate
3179 // uncommon trap or exception is thrown.
3180 Node* GraphKit::gen_checkcast(Node *obj, Node* superklass,
3181 Node* *failure_control) {
3182 kill_dead_locals(); // Benefit all the uncommon traps
3183 const TypeKlassPtr *tk = _gvn.type(superklass)->is_klassptr();
3184 const Type *toop = tk->cast_to_exactness(false)->as_instance_type();
3185
3186 // Fast cutout: Check the case that the cast is vacuously true.
3187 // This detects the common cases where the test will short-circuit
3188 // away completely. We do this before we perform the null check,
3189 // because if the test is going to turn into zero code, we don't
3190 // want a residual null check left around. (Causes a slowdown,
3191 // for example, in some objArray manipulations, such as a[i]=a[j].)
3192 if (tk->singleton()) {
3193 const TypeOopPtr* objtp = _gvn.type(obj)->isa_oopptr();
3194 if (objtp != NULL) {
3195 switch (C->static_subtype_check(tk, objtp->as_klass_type())) {
3196 case Compile::SSC_always_true:
3197 // If we know the type check always succeed then we don't use
3198 // the profiling data at this bytecode. Don't lose it, feed it
3199 // to the type system as a speculative type.
3200 return record_profiled_receiver_for_speculation(obj);
3201 case Compile::SSC_always_false:
3202 // It needs a null check because a null will *pass* the cast check.
3203 // A non-null value will always produce an exception.
3204 if (!objtp->maybe_null()) {
3205 bool is_aastore = (java_bc() == Bytecodes::_aastore);
3206 Deoptimization::DeoptReason reason = is_aastore ?
3207 Deoptimization::Reason_array_check : Deoptimization::Reason_class_check;
3208 builtin_throw(reason);
3209 return top();
3210 } else if (!too_many_traps_or_recompiles(Deoptimization::Reason_null_assert)) {
3211 return null_assert(obj);
3212 }
3213 break; // Fall through to full check
3214 default:
3215 break;
3216 }
3217 }
3218 }
3219
3220 ciProfileData* data = NULL;
3221 bool safe_for_replace = false;
3222 if (failure_control == NULL) { // use MDO in regular case only
3223 assert(java_bc() == Bytecodes::_aastore ||
3224 java_bc() == Bytecodes::_checkcast,
3225 "interpreter profiles type checks only for these BCs");
3226 data = method()->method_data()->bci_to_data(bci());
3227 safe_for_replace = true;
3228 }
3229
3230 // Make the merge point
3231 enum { _obj_path = 1, _null_path, PATH_LIMIT };
3232 RegionNode* region = new RegionNode(PATH_LIMIT);
3233 Node* phi = new PhiNode(region, toop);
3234 C->set_has_split_ifs(true); // Has chance for split-if optimization
3235
3236 // Use null-cast information if it is available
3237 bool speculative_not_null = false;
3238 bool never_see_null = ((failure_control == NULL) // regular case only
3239 && seems_never_null(obj, data, speculative_not_null));
3240
3241 // Null check; get casted pointer; set region slot 3
3242 Node* null_ctl = top();
3243 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null);
3244
3245 // If not_null_obj is dead, only null-path is taken
3246 if (stopped()) { // Doing instance-of on a NULL?
3247 set_control(null_ctl);
3248 return null();
3249 }
3250 region->init_req(_null_path, null_ctl);
3251 phi ->init_req(_null_path, null()); // Set null path value
3252 if (null_ctl == top()) {
3253 // Do this eagerly, so that pattern matches like is_diamond_phi
3254 // will work even during parsing.
3255 assert(_null_path == PATH_LIMIT-1, "delete last");
3256 region->del_req(_null_path);
3257 phi ->del_req(_null_path);
3258 }
3259
3260 Node* cast_obj = NULL;
3261 if (tk->klass_is_exact()) {
3262 // The following optimization tries to statically cast the speculative type of the object
3263 // (for example obtained during profiling) to the type of the superklass and then do a
3264 // dynamic check that the type of the object is what we expect. To work correctly
3265 // for checkcast and aastore the type of superklass should be exact.
3266 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3267 // We may not have profiling here or it may not help us. If we have
3268 // a speculative type use it to perform an exact cast.
3269 ciKlass* spec_obj_type = obj_type->speculative_type();
3270 if (spec_obj_type != NULL || data != NULL) {
3271 cast_obj = maybe_cast_profiled_receiver(not_null_obj, tk, spec_obj_type, safe_for_replace);
3272 if (cast_obj != NULL) {
3273 if (failure_control != NULL) // failure is now impossible
3274 (*failure_control) = top();
3275 // adjust the type of the phi to the exact klass:
3276 phi->raise_bottom_type(_gvn.type(cast_obj)->meet_speculative(TypePtr::NULL_PTR));
3277 }
3278 }
3279 }
3280
3281 if (cast_obj == NULL) {
3282 // Generate the subtype check
3283 Node* not_subtype_ctrl = gen_subtype_check(not_null_obj, superklass );
3284
3285 // Plug in success path into the merge
3286 cast_obj = _gvn.transform(new CheckCastPPNode(control(), not_null_obj, toop));
3287 // Failure path ends in uncommon trap (or may be dead - failure impossible)
3288 if (failure_control == NULL) {
3289 if (not_subtype_ctrl != top()) { // If failure is possible
3290 PreserveJVMState pjvms(this);
3291 set_control(not_subtype_ctrl);
3292 bool is_aastore = (java_bc() == Bytecodes::_aastore);
3293 Deoptimization::DeoptReason reason = is_aastore ?
3294 Deoptimization::Reason_array_check : Deoptimization::Reason_class_check;
3295 builtin_throw(reason);
3296 }
3297 } else {
3298 (*failure_control) = not_subtype_ctrl;
3299 }
3300 }
3301
3302 region->init_req(_obj_path, control());
3303 phi ->init_req(_obj_path, cast_obj);
3304
3305 // A merge of NULL or Casted-NotNull obj
3306 Node* res = _gvn.transform(phi);
3307
3308 // Note I do NOT always 'replace_in_map(obj,result)' here.
3309 // if( tk->klass()->can_be_primary_super() )
3310 // This means that if I successfully store an Object into an array-of-String
3311 // I 'forget' that the Object is really now known to be a String. I have to
3312 // do this because we don't have true union types for interfaces - if I store
3313 // a Baz into an array-of-Interface and then tell the optimizer it's an
3314 // Interface, I forget that it's also a Baz and cannot do Baz-like field
3315 // references to it. FIX THIS WHEN UNION TYPES APPEAR!
3316 // replace_in_map( obj, res );
3317
3318 // Return final merged results
3319 set_control( _gvn.transform(region) );
3320 record_for_igvn(region);
3321
3322 return record_profiled_receiver_for_speculation(res);
3323 }
3324
3325 //------------------------------next_monitor-----------------------------------
3326 // What number should be given to the next monitor?
3327 int GraphKit::next_monitor() {
3328 int current = jvms()->monitor_depth()* C->sync_stack_slots();
3329 int next = current + C->sync_stack_slots();
3330 // Keep the toplevel high water mark current:
3331 if (C->fixed_slots() < next) C->set_fixed_slots(next);
3332 return current;
3333 }
3334
3335 //------------------------------insert_mem_bar---------------------------------
3336 // Memory barrier to avoid floating things around
3337 // The membar serves as a pinch point between both control and all memory slices.
3338 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) {
3339 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent);
3340 mb->init_req(TypeFunc::Control, control());
3341 mb->init_req(TypeFunc::Memory, reset_memory());
3342 Node* membar = _gvn.transform(mb);
3370 }
3371 Node* membar = _gvn.transform(mb);
3372 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control)));
3373 if (alias_idx == Compile::AliasIdxBot) {
3374 merged_memory()->set_base_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)));
3375 } else {
3376 set_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)),alias_idx);
3377 }
3378 return membar;
3379 }
3380
3381 //------------------------------shared_lock------------------------------------
3382 // Emit locking code.
3383 FastLockNode* GraphKit::shared_lock(Node* obj) {
3384 // bci is either a monitorenter bc or InvocationEntryBci
3385 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
3386 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
3387
3388 if( !GenerateSynchronizationCode )
3389 return NULL; // Not locking things?
3390 if (stopped()) // Dead monitor?
3391 return NULL;
3392
3393 assert(dead_locals_are_killed(), "should kill locals before sync. point");
3394
3395 // Box the stack location
3396 Node* box = _gvn.transform(new BoxLockNode(next_monitor()));
3397 Node* mem = reset_memory();
3398
3399 FastLockNode * flock = _gvn.transform(new FastLockNode(0, obj, box) )->as_FastLock();
3400
3401 // Create the rtm counters for this fast lock if needed.
3402 flock->create_rtm_lock_counter(sync_jvms()); // sync_jvms used to get current bci
3403
3404 // Add monitor to debug info for the slow path. If we block inside the
3405 // slow path and de-opt, we need the monitor hanging around
3406 map()->push_monitor( flock );
3407
3408 const TypeFunc *tf = LockNode::lock_type();
3409 LockNode *lock = new LockNode(C, tf);
3438 }
3439 #endif
3440
3441 return flock;
3442 }
3443
3444
3445 //------------------------------shared_unlock----------------------------------
3446 // Emit unlocking code.
3447 void GraphKit::shared_unlock(Node* box, Node* obj) {
3448 // bci is either a monitorenter bc or InvocationEntryBci
3449 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
3450 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
3451
3452 if( !GenerateSynchronizationCode )
3453 return;
3454 if (stopped()) { // Dead monitor?
3455 map()->pop_monitor(); // Kill monitor from debug info
3456 return;
3457 }
3458
3459 // Memory barrier to avoid floating things down past the locked region
3460 insert_mem_bar(Op_MemBarReleaseLock);
3461
3462 const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type();
3463 UnlockNode *unlock = new UnlockNode(C, tf);
3464 #ifdef ASSERT
3465 unlock->set_dbg_jvms(sync_jvms());
3466 #endif
3467 uint raw_idx = Compile::AliasIdxRaw;
3468 unlock->init_req( TypeFunc::Control, control() );
3469 unlock->init_req( TypeFunc::Memory , memory(raw_idx) );
3470 unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
3471 unlock->init_req( TypeFunc::FramePtr, frameptr() );
3472 unlock->init_req( TypeFunc::ReturnAdr, top() );
3473
3474 unlock->init_req(TypeFunc::Parms + 0, obj);
3475 unlock->init_req(TypeFunc::Parms + 1, box);
3476 unlock = _gvn.transform(unlock)->as_Unlock();
3477
3478 Node* mem = reset_memory();
3479
3480 // unlock has no side-effects, sets few values
3481 set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM);
3482
3483 // Kill monitor from debug info
3484 map()->pop_monitor( );
3485 }
3486
3487 //-------------------------------get_layout_helper-----------------------------
3488 // If the given klass is a constant or known to be an array,
3489 // fetch the constant layout helper value into constant_value
3490 // and return (Node*)NULL. Otherwise, load the non-constant
3491 // layout helper value, and return the node which represents it.
3492 // This two-faced routine is useful because allocation sites
3493 // almost always feature constant types.
3494 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) {
3495 const TypeKlassPtr* inst_klass = _gvn.type(klass_node)->isa_klassptr();
3496 if (!StressReflectiveCode && inst_klass != NULL) {
3497 bool xklass = inst_klass->klass_is_exact();
3498 if (xklass || inst_klass->isa_aryklassptr()) {
3499 jint lhelper;
3500 if (inst_klass->isa_aryklassptr()) {
3501 BasicType elem = inst_klass->as_instance_type()->isa_aryptr()->elem()->array_element_basic_type();
3502 if (is_reference_type(elem, true)) {
3503 elem = T_OBJECT;
3504 }
3505 lhelper = Klass::array_layout_helper(elem);
3506 } else {
3507 lhelper = inst_klass->is_instklassptr()->exact_klass()->layout_helper();
3508 }
3509 if (lhelper != Klass::_lh_neutral_value) {
3510 constant_value = lhelper;
3511 return (Node*) NULL;
3512 }
3513 }
3514 }
3515 constant_value = Klass::_lh_neutral_value; // put in a known value
3516 Node* lhp = basic_plus_adr(klass_node, klass_node, in_bytes(Klass::layout_helper_offset()));
3517 return make_load(NULL, lhp, TypeInt::INT, T_INT, MemNode::unordered);
3518 }
3519
3520 // We just put in an allocate/initialize with a big raw-memory effect.
3521 // Hook selected additional alias categories on the initialization.
3522 static void hook_memory_on_init(GraphKit& kit, int alias_idx,
3523 MergeMemNode* init_in_merge,
3524 Node* init_out_raw) {
3525 DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory());
3526 assert(init_in_merge->memory_at(alias_idx) == init_in_raw, "");
3527
3528 Node* prevmem = kit.memory(alias_idx);
3529 init_in_merge->set_memory_at(alias_idx, prevmem);
3530 kit.set_memory(init_out_raw, alias_idx);
3531 }
3532
3533 //---------------------------set_output_for_allocation-------------------------
3534 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc,
3535 const TypeOopPtr* oop_type,
3536 bool deoptimize_on_exception) {
3537 int rawidx = Compile::AliasIdxRaw;
3538 alloc->set_req( TypeFunc::FramePtr, frameptr() );
3539 add_safepoint_edges(alloc);
3540 Node* allocx = _gvn.transform(alloc);
3541 set_control( _gvn.transform(new ProjNode(allocx, TypeFunc::Control) ) );
3542 // create memory projection for i_o
3543 set_memory ( _gvn.transform( new ProjNode(allocx, TypeFunc::Memory, true) ), rawidx );
3544 make_slow_call_ex(allocx, env()->Throwable_klass(), true, deoptimize_on_exception);
3545
3546 // create a memory projection as for the normal control path
3547 Node* malloc = _gvn.transform(new ProjNode(allocx, TypeFunc::Memory));
3548 set_memory(malloc, rawidx);
3549
3550 // a normal slow-call doesn't change i_o, but an allocation does
3551 // we create a separate i_o projection for the normal control path
3552 set_i_o(_gvn.transform( new ProjNode(allocx, TypeFunc::I_O, false) ) );
3553 Node* rawoop = _gvn.transform( new ProjNode(allocx, TypeFunc::Parms) );
3554
3555 // put in an initialization barrier
3556 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx,
3557 rawoop)->as_Initialize();
3558 assert(alloc->initialization() == init, "2-way macro link must work");
3559 assert(init ->allocation() == alloc, "2-way macro link must work");
3560 {
3561 // Extract memory strands which may participate in the new object's
3562 // initialization, and source them from the new InitializeNode.
3563 // This will allow us to observe initializations when they occur,
3564 // and link them properly (as a group) to the InitializeNode.
3565 assert(init->in(InitializeNode::Memory) == malloc, "");
3566 MergeMemNode* minit_in = MergeMemNode::make(malloc);
3567 init->set_req(InitializeNode::Memory, minit_in);
3568 record_for_igvn(minit_in); // fold it up later, if possible
3569 Node* minit_out = memory(rawidx);
3570 assert(minit_out->is_Proj() && minit_out->in(0) == init, "");
3571 // Add an edge in the MergeMem for the header fields so an access
3572 // to one of those has correct memory state
3573 set_memory(minit_out, C->get_alias_index(oop_type->add_offset(oopDesc::mark_offset_in_bytes())));
3574 set_memory(minit_out, C->get_alias_index(oop_type->add_offset(oopDesc::klass_offset_in_bytes())));
3575 if (oop_type->isa_aryptr()) {
3576 const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot);
3577 int elemidx = C->get_alias_index(telemref);
3578 hook_memory_on_init(*this, elemidx, minit_in, minit_out);
3579 } else if (oop_type->isa_instptr()) {
3580 ciInstanceKlass* ik = oop_type->is_instptr()->instance_klass();
3581 for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) {
3582 ciField* field = ik->nonstatic_field_at(i);
3583 if (field->offset() >= TrackedInitializationLimit * HeapWordSize)
3584 continue; // do not bother to track really large numbers of fields
3585 // Find (or create) the alias category for this field:
3586 int fieldidx = C->alias_type(field)->index();
3587 hook_memory_on_init(*this, fieldidx, minit_in, minit_out);
3588 }
3589 }
3590 }
3591
3592 // Cast raw oop to the real thing...
3593 Node* javaoop = new CheckCastPPNode(control(), rawoop, oop_type);
3594 javaoop = _gvn.transform(javaoop);
3595 C->set_recent_alloc(control(), javaoop);
3596 assert(just_allocated_object(control()) == javaoop, "just allocated");
3597
3598 #ifdef ASSERT
3599 { // Verify that the AllocateNode::Ideal_allocation recognizers work:
3610 assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please");
3611 }
3612 }
3613 #endif //ASSERT
3614
3615 return javaoop;
3616 }
3617
3618 //---------------------------new_instance--------------------------------------
3619 // This routine takes a klass_node which may be constant (for a static type)
3620 // or may be non-constant (for reflective code). It will work equally well
3621 // for either, and the graph will fold nicely if the optimizer later reduces
3622 // the type to a constant.
3623 // The optional arguments are for specialized use by intrinsics:
3624 // - If 'extra_slow_test' if not null is an extra condition for the slow-path.
3625 // - If 'return_size_val', report the total object size to the caller.
3626 // - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize)
3627 Node* GraphKit::new_instance(Node* klass_node,
3628 Node* extra_slow_test,
3629 Node* *return_size_val,
3630 bool deoptimize_on_exception) {
3631 // Compute size in doublewords
3632 // The size is always an integral number of doublewords, represented
3633 // as a positive bytewise size stored in the klass's layout_helper.
3634 // The layout_helper also encodes (in a low bit) the need for a slow path.
3635 jint layout_con = Klass::_lh_neutral_value;
3636 Node* layout_val = get_layout_helper(klass_node, layout_con);
3637 int layout_is_con = (layout_val == NULL);
3638
3639 if (extra_slow_test == NULL) extra_slow_test = intcon(0);
3640 // Generate the initial go-slow test. It's either ALWAYS (return a
3641 // Node for 1) or NEVER (return a NULL) or perhaps (in the reflective
3642 // case) a computed value derived from the layout_helper.
3643 Node* initial_slow_test = NULL;
3644 if (layout_is_con) {
3645 assert(!StressReflectiveCode, "stress mode does not use these paths");
3646 bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con);
3647 initial_slow_test = must_go_slow ? intcon(1) : extra_slow_test;
3648 } else { // reflective case
3649 // This reflective path is used by Unsafe.allocateInstance.
3650 // (It may be stress-tested by specifying StressReflectiveCode.)
3651 // Basically, we want to get into the VM is there's an illegal argument.
3652 Node* bit = intcon(Klass::_lh_instance_slow_path_bit);
3653 initial_slow_test = _gvn.transform( new AndINode(layout_val, bit) );
3654 if (extra_slow_test != intcon(0)) {
3655 initial_slow_test = _gvn.transform( new OrINode(initial_slow_test, extra_slow_test) );
3656 }
3657 // (Macro-expander will further convert this to a Bool, if necessary.)
3668
3669 // Clear the low bits to extract layout_helper_size_in_bytes:
3670 assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit");
3671 Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong));
3672 size = _gvn.transform( new AndXNode(size, mask) );
3673 }
3674 if (return_size_val != NULL) {
3675 (*return_size_val) = size;
3676 }
3677
3678 // This is a precise notnull oop of the klass.
3679 // (Actually, it need not be precise if this is a reflective allocation.)
3680 // It's what we cast the result to.
3681 const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr();
3682 if (!tklass) tklass = TypeInstKlassPtr::OBJECT;
3683 const TypeOopPtr* oop_type = tklass->as_instance_type();
3684
3685 // Now generate allocation code
3686
3687 // The entire memory state is needed for slow path of the allocation
3688 // since GC and deoptimization can happened.
3689 Node *mem = reset_memory();
3690 set_all_memory(mem); // Create new memory state
3691
3692 AllocateNode* alloc = new AllocateNode(C, AllocateNode::alloc_type(Type::TOP),
3693 control(), mem, i_o(),
3694 size, klass_node,
3695 initial_slow_test);
3696
3697 return set_output_for_allocation(alloc, oop_type, deoptimize_on_exception);
3698 }
3699
3700 //-------------------------------new_array-------------------------------------
3701 // helper for both newarray and anewarray
3702 // The 'length' parameter is (obviously) the length of the array.
3703 // See comments on new_instance for the meaning of the other arguments.
3704 Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable)
3705 Node* length, // number of array elements
3706 int nargs, // number of arguments to push back for uncommon trap
3707 Node* *return_size_val,
3708 bool deoptimize_on_exception) {
3709 jint layout_con = Klass::_lh_neutral_value;
3710 Node* layout_val = get_layout_helper(klass_node, layout_con);
3711 int layout_is_con = (layout_val == NULL);
3712
3713 if (!layout_is_con && !StressReflectiveCode &&
3714 !too_many_traps(Deoptimization::Reason_class_check)) {
3715 // This is a reflective array creation site.
3716 // Optimistically assume that it is a subtype of Object[],
3717 // so that we can fold up all the address arithmetic.
3718 layout_con = Klass::array_layout_helper(T_OBJECT);
3719 Node* cmp_lh = _gvn.transform( new CmpINode(layout_val, intcon(layout_con)) );
3720 Node* bol_lh = _gvn.transform( new BoolNode(cmp_lh, BoolTest::eq) );
3721 { BuildCutout unless(this, bol_lh, PROB_MAX);
3722 inc_sp(nargs);
3723 uncommon_trap(Deoptimization::Reason_class_check,
3724 Deoptimization::Action_maybe_recompile);
3725 }
3726 layout_val = NULL;
3727 layout_is_con = true;
3728 }
3729
3730 // Generate the initial go-slow test. Make sure we do not overflow
3731 // if length is huge (near 2Gig) or negative! We do not need
3732 // exact double-words here, just a close approximation of needed
3733 // double-words. We can't add any offset or rounding bits, lest we
3734 // take a size -1 of bytes and make it positive. Use an unsigned
3735 // compare, so negative sizes look hugely positive.
3736 int fast_size_limit = FastAllocateSizeLimit;
3737 if (layout_is_con) {
3738 assert(!StressReflectiveCode, "stress mode does not use these paths");
3739 // Increase the size limit if we have exact knowledge of array type.
3740 int log2_esize = Klass::layout_helper_log2_element_size(layout_con);
3741 fast_size_limit <<= (LogBytesPerLong - log2_esize);
3742 }
3743
3744 Node* initial_slow_cmp = _gvn.transform( new CmpUNode( length, intcon( fast_size_limit ) ) );
3745 Node* initial_slow_test = _gvn.transform( new BoolNode( initial_slow_cmp, BoolTest::gt ) );
3746
3747 // --- Size Computation ---
3748 // array_size = round_to_heap(array_header + (length << elem_shift));
3749 // where round_to_heap(x) == align_to(x, MinObjAlignmentInBytes)
3750 // and align_to(x, y) == ((x + y-1) & ~(y-1))
3751 // The rounding mask is strength-reduced, if possible.
3752 int round_mask = MinObjAlignmentInBytes - 1;
3753 Node* header_size = NULL;
3754 int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
3755 // (T_BYTE has the weakest alignment and size restrictions...)
3756 if (layout_is_con) {
3757 int hsize = Klass::layout_helper_header_size(layout_con);
3758 int eshift = Klass::layout_helper_log2_element_size(layout_con);
3759 BasicType etype = Klass::layout_helper_element_type(layout_con);
3760 if ((round_mask & ~right_n_bits(eshift)) == 0)
3761 round_mask = 0; // strength-reduce it if it goes away completely
3762 assert((hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded");
3763 assert(header_size_min <= hsize, "generic minimum is smallest");
3764 header_size_min = hsize;
3765 header_size = intcon(hsize + round_mask);
3766 } else {
3767 Node* hss = intcon(Klass::_lh_header_size_shift);
3768 Node* hsm = intcon(Klass::_lh_header_size_mask);
3769 Node* hsize = _gvn.transform( new URShiftINode(layout_val, hss) );
3770 hsize = _gvn.transform( new AndINode(hsize, hsm) );
3771 Node* mask = intcon(round_mask);
3772 header_size = _gvn.transform( new AddINode(hsize, mask) );
3773 }
3774
3775 Node* elem_shift = NULL;
3776 if (layout_is_con) {
3777 int eshift = Klass::layout_helper_log2_element_size(layout_con);
3778 if (eshift != 0)
3779 elem_shift = intcon(eshift);
3780 } else {
3781 // There is no need to mask or shift this value.
3782 // The semantics of LShiftINode include an implicit mask to 0x1F.
3826 // places, one where the length is sharply limited, and the other
3827 // after a successful allocation.
3828 Node* abody = lengthx;
3829 if (elem_shift != NULL)
3830 abody = _gvn.transform( new LShiftXNode(lengthx, elem_shift) );
3831 Node* size = _gvn.transform( new AddXNode(headerx, abody) );
3832 if (round_mask != 0) {
3833 Node* mask = MakeConX(~round_mask);
3834 size = _gvn.transform( new AndXNode(size, mask) );
3835 }
3836 // else if round_mask == 0, the size computation is self-rounding
3837
3838 if (return_size_val != NULL) {
3839 // This is the size
3840 (*return_size_val) = size;
3841 }
3842
3843 // Now generate allocation code
3844
3845 // The entire memory state is needed for slow path of the allocation
3846 // since GC and deoptimization can happened.
3847 Node *mem = reset_memory();
3848 set_all_memory(mem); // Create new memory state
3849
3850 if (initial_slow_test->is_Bool()) {
3851 // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick.
3852 initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn);
3853 }
3854
3855 // Create the AllocateArrayNode and its result projections
3856 AllocateArrayNode* alloc
3857 = new AllocateArrayNode(C, AllocateArrayNode::alloc_type(TypeInt::INT),
3858 control(), mem, i_o(),
3859 size, klass_node,
3860 initial_slow_test,
3861 length);
3862
3863 // Cast to correct type. Note that the klass_node may be constant or not,
3864 // and in the latter case the actual array type will be inexact also.
3865 // (This happens via a non-constant argument to inline_native_newArray.)
3866 // In any case, the value of klass_node provides the desired array type.
3867 const TypeInt* length_type = _gvn.find_int_type(length);
3868 const TypeOopPtr* ary_type = _gvn.type(klass_node)->is_klassptr()->as_instance_type();
3869 if (ary_type->isa_aryptr() && length_type != NULL) {
3870 // Try to get a better type than POS for the size
3871 ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
3872 }
3873
3874 Node* javaoop = set_output_for_allocation(alloc, ary_type, deoptimize_on_exception);
3875
3876 array_ideal_length(alloc, ary_type, true);
3877 return javaoop;
3878 }
3879
3880 // The following "Ideal_foo" functions are placed here because they recognize
3881 // the graph shapes created by the functions immediately above.
3882
3883 //---------------------------Ideal_allocation----------------------------------
3884 // Given an oop pointer or raw pointer, see if it feeds from an AllocateNode.
3885 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase) {
3886 if (ptr == NULL) { // reduce dumb test in callers
3887 return NULL;
3888 }
3997 set_all_memory(ideal.merged_memory());
3998 set_i_o(ideal.i_o());
3999 set_control(ideal.ctrl());
4000 }
4001
4002 void GraphKit::final_sync(IdealKit& ideal) {
4003 // Final sync IdealKit and graphKit.
4004 sync_kit(ideal);
4005 }
4006
4007 Node* GraphKit::load_String_length(Node* str, bool set_ctrl) {
4008 Node* len = load_array_length(load_String_value(str, set_ctrl));
4009 Node* coder = load_String_coder(str, set_ctrl);
4010 // Divide length by 2 if coder is UTF16
4011 return _gvn.transform(new RShiftINode(len, coder));
4012 }
4013
4014 Node* GraphKit::load_String_value(Node* str, bool set_ctrl) {
4015 int value_offset = java_lang_String::value_offset();
4016 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4017 false, NULL, 0);
4018 const TypePtr* value_field_type = string_type->add_offset(value_offset);
4019 const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::NotNull,
4020 TypeAry::make(TypeInt::BYTE, TypeInt::POS),
4021 ciTypeArrayKlass::make(T_BYTE), true, 0);
4022 Node* p = basic_plus_adr(str, str, value_offset);
4023 Node* load = access_load_at(str, p, value_field_type, value_type, T_OBJECT,
4024 IN_HEAP | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0) | MO_UNORDERED);
4025 return load;
4026 }
4027
4028 Node* GraphKit::load_String_coder(Node* str, bool set_ctrl) {
4029 if (!CompactStrings) {
4030 return intcon(java_lang_String::CODER_UTF16);
4031 }
4032 int coder_offset = java_lang_String::coder_offset();
4033 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4034 false, NULL, 0);
4035 const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
4036
4037 Node* p = basic_plus_adr(str, str, coder_offset);
4038 Node* load = access_load_at(str, p, coder_field_type, TypeInt::BYTE, T_BYTE,
4039 IN_HEAP | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0) | MO_UNORDERED);
4040 return load;
4041 }
4042
4043 void GraphKit::store_String_value(Node* str, Node* value) {
4044 int value_offset = java_lang_String::value_offset();
4045 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4046 false, NULL, 0);
4047 const TypePtr* value_field_type = string_type->add_offset(value_offset);
4048
4049 access_store_at(str, basic_plus_adr(str, value_offset), value_field_type,
4050 value, TypeAryPtr::BYTES, T_OBJECT, IN_HEAP | MO_UNORDERED);
4051 }
4052
4053 void GraphKit::store_String_coder(Node* str, Node* value) {
4054 int coder_offset = java_lang_String::coder_offset();
4055 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4056 false, NULL, 0);
4057 const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
4058
4059 access_store_at(str, basic_plus_adr(str, coder_offset), coder_field_type,
4060 value, TypeInt::BYTE, T_BYTE, IN_HEAP | MO_UNORDERED);
4061 }
4062
4063 // Capture src and dst memory state with a MergeMemNode
4064 Node* GraphKit::capture_memory(const TypePtr* src_type, const TypePtr* dst_type) {
4065 if (src_type == dst_type) {
4066 // Types are equal, we don't need a MergeMemNode
4067 return memory(src_type);
4068 }
4069 MergeMemNode* merge = MergeMemNode::make(map()->memory());
4070 record_for_igvn(merge); // fold it up later, if possible
4071 int src_idx = C->get_alias_index(src_type);
4072 int dst_idx = C->get_alias_index(dst_type);
4073 merge->set_memory_at(src_idx, memory(src_idx));
4074 merge->set_memory_at(dst_idx, memory(dst_idx));
4075 return merge;
4076 }
4149 i_char->init_req(2, AddI(i_char, intcon(2)));
4150
4151 set_control(IfFalse(iff));
4152 set_memory(st, TypeAryPtr::BYTES);
4153 }
4154
4155 Node* GraphKit::make_constant_from_field(ciField* field, Node* obj) {
4156 if (!field->is_constant()) {
4157 return NULL; // Field not marked as constant.
4158 }
4159 ciInstance* holder = NULL;
4160 if (!field->is_static()) {
4161 ciObject* const_oop = obj->bottom_type()->is_oopptr()->const_oop();
4162 if (const_oop != NULL && const_oop->is_instance()) {
4163 holder = const_oop->as_instance();
4164 }
4165 }
4166 const Type* con_type = Type::make_constant_from_field(field, holder, field->layout_type(),
4167 /*is_unsigned_load=*/false);
4168 if (con_type != NULL) {
4169 return makecon(con_type);
4170 }
4171 return NULL;
4172 }
|
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "ci/ciFlatArrayKlass.hpp"
27 #include "ci/ciInlineKlass.hpp"
28 #include "ci/ciUtilities.hpp"
29 #include "classfile/javaClasses.hpp"
30 #include "ci/ciObjArray.hpp"
31 #include "asm/register.hpp"
32 #include "compiler/compileLog.hpp"
33 #include "gc/shared/barrierSet.hpp"
34 #include "gc/shared/c2/barrierSetC2.hpp"
35 #include "interpreter/interpreter.hpp"
36 #include "memory/resourceArea.hpp"
37 #include "opto/addnode.hpp"
38 #include "opto/castnode.hpp"
39 #include "opto/convertnode.hpp"
40 #include "opto/graphKit.hpp"
41 #include "opto/idealKit.hpp"
42 #include "opto/inlinetypenode.hpp"
43 #include "opto/intrinsicnode.hpp"
44 #include "opto/locknode.hpp"
45 #include "opto/machnode.hpp"
46 #include "opto/narrowptrnode.hpp"
47 #include "opto/opaquenode.hpp"
48 #include "opto/parse.hpp"
49 #include "opto/rootnode.hpp"
50 #include "opto/runtime.hpp"
51 #include "opto/subtypenode.hpp"
52 #include "runtime/deoptimization.hpp"
53 #include "runtime/sharedRuntime.hpp"
54 #include "utilities/bitMap.inline.hpp"
55 #include "utilities/powerOfTwo.hpp"
56 #include "utilities/growableArray.hpp"
57
58 //----------------------------GraphKit-----------------------------------------
59 // Main utility constructor.
60 GraphKit::GraphKit(JVMState* jvms, PhaseGVN* gvn)
61 : Phase(Phase::Parser),
62 _env(C->env()),
63 _gvn((gvn != NULL) ? *gvn : *C->initial_gvn()),
64 _barrier_set(BarrierSet::barrier_set()->barrier_set_c2())
65 {
66 assert(gvn == NULL || !gvn->is_IterGVN() || gvn->is_IterGVN()->delay_transform(), "delay transform should be enabled");
67 _exceptions = jvms->map()->next_exception();
68 if (_exceptions != NULL) jvms->map()->set_next_exception(NULL);
69 set_jvms(jvms);
70 #ifdef ASSERT
71 if (_gvn.is_IterGVN() != NULL) {
72 assert(_gvn.is_IterGVN()->delay_transform(), "Transformation must be delayed if IterGVN is used");
73 // Save the initial size of _for_igvn worklist for verification (see ~GraphKit)
74 _worklist_size = _gvn.C->for_igvn()->size();
75 }
76 #endif
77 }
78
79 // Private constructor for parser.
80 GraphKit::GraphKit()
81 : Phase(Phase::Parser),
82 _env(C->env()),
83 _gvn(*C->initial_gvn()),
84 _barrier_set(BarrierSet::barrier_set()->barrier_set_c2())
85 {
86 _exceptions = NULL;
87 set_map(NULL);
88 debug_only(_sp = -99);
89 debug_only(set_bci(-99));
90 }
91
92
93
94 //---------------------------clean_stack---------------------------------------
95 // Clear away rubbish from the stack area of the JVM state.
96 // This destroys any arguments that may be waiting on the stack.
829 if (PrintMiscellaneous && (Verbose || WizardMode)) {
830 tty->print_cr("Zombie local %d: ", local);
831 jvms->dump();
832 }
833 return false;
834 }
835 }
836 }
837 return true;
838 }
839
840 #endif //ASSERT
841
842 // Helper function for enforcing certain bytecodes to reexecute if deoptimization happens.
843 static bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) {
844 ciMethod* cur_method = jvms->method();
845 int cur_bci = jvms->bci();
846 if (cur_method != NULL && cur_bci != InvocationEntryBci) {
847 Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci);
848 return Interpreter::bytecode_should_reexecute(code) ||
849 (is_anewarray && (code == Bytecodes::_multianewarray));
850 // Reexecute _multianewarray bytecode which was replaced with
851 // sequence of [a]newarray. See Parse::do_multianewarray().
852 //
853 // Note: interpreter should not have it set since this optimization
854 // is limited by dimensions and guarded by flag so in some cases
855 // multianewarray() runtime calls will be generated and
856 // the bytecode should not be reexecutes (stack will not be reset).
857 } else {
858 return false;
859 }
860 }
861
862 // Helper function for adding JVMState and debug information to node
863 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) {
864 // Add the safepoint edges to the call (or other safepoint).
865
866 // Make sure dead locals are set to top. This
867 // should help register allocation time and cut down on the size
868 // of the deoptimization information.
869 assert(dead_locals_are_killed(), "garbage in debug info before safepoint");
1089 ciSignature* declared_signature = NULL;
1090 ciMethod* ignored_callee = method()->get_method_at_bci(bci(), ignored_will_link, &declared_signature);
1091 assert(declared_signature != NULL, "cannot be null");
1092 inputs = declared_signature->arg_size_for_bc(code);
1093 int size = declared_signature->return_type()->size();
1094 depth = size - inputs;
1095 }
1096 break;
1097
1098 case Bytecodes::_multianewarray:
1099 {
1100 ciBytecodeStream iter(method());
1101 iter.reset_to_bci(bci());
1102 iter.next();
1103 inputs = iter.get_dimensions();
1104 assert(rsize == 1, "");
1105 depth = rsize - inputs;
1106 }
1107 break;
1108
1109 case Bytecodes::_withfield: {
1110 bool ignored_will_link;
1111 ciField* field = method()->get_field_at_bci(bci(), ignored_will_link);
1112 int size = field->type()->size();
1113 inputs = size+1;
1114 depth = rsize - inputs;
1115 break;
1116 }
1117
1118 case Bytecodes::_ireturn:
1119 case Bytecodes::_lreturn:
1120 case Bytecodes::_freturn:
1121 case Bytecodes::_dreturn:
1122 case Bytecodes::_areturn:
1123 assert(rsize == -depth, "");
1124 inputs = rsize;
1125 break;
1126
1127 case Bytecodes::_jsr:
1128 case Bytecodes::_jsr_w:
1129 inputs = 0;
1130 depth = 1; // S.B. depth=1, not zero
1131 break;
1132
1133 default:
1134 // bytecode produces a typed result
1135 inputs = rsize - depth;
1136 assert(inputs >= 0, "");
1137 break;
1180 Node* conv = _gvn.transform( new ConvI2LNode(offset));
1181 Node* mask = _gvn.transform(ConLNode::make((julong) max_juint));
1182 return _gvn.transform( new AndLNode(conv, mask) );
1183 }
1184
1185 Node* GraphKit::ConvL2I(Node* offset) {
1186 // short-circuit a common case
1187 jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot);
1188 if (offset_con != (jlong)Type::OffsetBot) {
1189 return intcon((int) offset_con);
1190 }
1191 return _gvn.transform( new ConvL2INode(offset));
1192 }
1193
1194 //-------------------------load_object_klass-----------------------------------
1195 Node* GraphKit::load_object_klass(Node* obj) {
1196 // Special-case a fresh allocation to avoid building nodes:
1197 Node* akls = AllocateNode::Ideal_klass(obj, &_gvn);
1198 if (akls != NULL) return akls;
1199 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
1200 return _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), k_adr, TypeInstPtr::KLASS, TypeInstKlassPtr::OBJECT));
1201 }
1202
1203 //-------------------------load_array_length-----------------------------------
1204 Node* GraphKit::load_array_length(Node* array) {
1205 // Special-case a fresh allocation to avoid building nodes:
1206 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array, &_gvn);
1207 Node *alen;
1208 if (alloc == NULL) {
1209 Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes());
1210 alen = _gvn.transform( new LoadRangeNode(0, immutable_memory(), r_adr, TypeInt::POS));
1211 } else {
1212 alen = array_ideal_length(alloc, _gvn.type(array)->is_oopptr(), false);
1213 }
1214 return alen;
1215 }
1216
1217 Node* GraphKit::array_ideal_length(AllocateArrayNode* alloc,
1218 const TypeOopPtr* oop_type,
1219 bool replace_length_in_map) {
1220 Node* length = alloc->Ideal_length();
1229 replace_in_map(length, ccast);
1230 }
1231 return ccast;
1232 }
1233 }
1234 return length;
1235 }
1236
1237 //------------------------------do_null_check----------------------------------
1238 // Helper function to do a NULL pointer check. Returned value is
1239 // the incoming address with NULL casted away. You are allowed to use the
1240 // not-null value only if you are control dependent on the test.
1241 #ifndef PRODUCT
1242 extern int explicit_null_checks_inserted,
1243 explicit_null_checks_elided;
1244 #endif
1245 Node* GraphKit::null_check_common(Node* value, BasicType type,
1246 // optional arguments for variations:
1247 bool assert_null,
1248 Node* *null_control,
1249 bool speculative,
1250 bool is_init_check) {
1251 assert(!assert_null || null_control == NULL, "not both at once");
1252 if (stopped()) return top();
1253 NOT_PRODUCT(explicit_null_checks_inserted++);
1254
1255 if (value->is_InlineType()) {
1256 InlineTypeNode* vt = value->as_InlineType();
1257 null_check_common(vt->get_is_init(), T_INT, assert_null, null_control, speculative, true);
1258 if (stopped()) {
1259 return top();
1260 }
1261 if (assert_null) {
1262 // TODO 8284443 Scalarize here (this currently leads to compilation bailouts)
1263 // vt = InlineTypeNode::make_null(_gvn, vt->type()->inline_klass());
1264 // replace_in_map(value, vt);
1265 // return vt;
1266 return null();
1267 }
1268 bool do_replace_in_map = (null_control == NULL || (*null_control) == top());
1269 return cast_not_null(value, do_replace_in_map);
1270 } else if (value->is_InlineTypePtr()) {
1271 // Null checking a scalarized but nullable inline type. Check the IsInit
1272 // input instead of the oop input to avoid keeping buffer allocations alive.
1273 InlineTypePtrNode* vtptr = value->as_InlineTypePtr();
1274 while (vtptr->get_oop()->is_InlineTypePtr()) {
1275 vtptr = vtptr->get_oop()->as_InlineTypePtr();
1276 }
1277 null_check_common(vtptr->get_is_init(), T_INT, assert_null, null_control, speculative, true);
1278 if (stopped()) {
1279 return top();
1280 }
1281 if (assert_null) {
1282 // TODO 8284443 Scalarize here (this currently leads to compilation bailouts)
1283 // vtptr = InlineTypePtrNode::make_null(_gvn, vtptr->type()->inline_klass());
1284 // replace_in_map(value, vtptr);
1285 // return vtptr;
1286 return null();
1287 }
1288 bool do_replace_in_map = (null_control == NULL || (*null_control) == top());
1289 return cast_not_null(value, do_replace_in_map);
1290 }
1291
1292 // Construct NULL check
1293 Node *chk = NULL;
1294 switch(type) {
1295 case T_LONG : chk = new CmpLNode(value, _gvn.zerocon(T_LONG)); break;
1296 case T_INT : chk = new CmpINode(value, _gvn.intcon(0)); break;
1297 case T_PRIMITIVE_OBJECT : // fall through
1298 case T_ARRAY : // fall through
1299 type = T_OBJECT; // simplify further tests
1300 case T_OBJECT : {
1301 const Type *t = _gvn.type( value );
1302
1303 const TypeOopPtr* tp = t->isa_oopptr();
1304 if (tp != NULL && !tp->is_loaded()
1305 // Only for do_null_check, not any of its siblings:
1306 && !assert_null && null_control == NULL) {
1307 // Usually, any field access or invocation on an unloaded oop type
1308 // will simply fail to link, since the statically linked class is
1309 // likely also to be unloaded. However, in -Xcomp mode, sometimes
1310 // the static class is loaded but the sharper oop type is not.
1311 // Rather than checking for this obscure case in lots of places,
1312 // we simply observe that a null check on an unloaded class
1313 // will always be followed by a nonsense operation, so we
1314 // can just issue the uncommon trap here.
1315 // Our access to the unloaded class will only be correct
1316 // after it has been loaded and initialized, which requires
1317 // a trip through the interpreter.
1376 }
1377 Node *oldcontrol = control();
1378 set_control(cfg);
1379 Node *res = cast_not_null(value);
1380 set_control(oldcontrol);
1381 NOT_PRODUCT(explicit_null_checks_elided++);
1382 return res;
1383 }
1384 cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true);
1385 if (cfg == NULL) break; // Quit at region nodes
1386 depth++;
1387 }
1388 }
1389
1390 //-----------
1391 // Branch to failure if null
1392 float ok_prob = PROB_MAX; // a priori estimate: nulls never happen
1393 Deoptimization::DeoptReason reason;
1394 if (assert_null) {
1395 reason = Deoptimization::reason_null_assert(speculative);
1396 } else if (type == T_OBJECT || is_init_check) {
1397 reason = Deoptimization::reason_null_check(speculative);
1398 } else {
1399 reason = Deoptimization::Reason_div0_check;
1400 }
1401 // %%% Since Reason_unhandled is not recorded on a per-bytecode basis,
1402 // ciMethodData::has_trap_at will return a conservative -1 if any
1403 // must-be-null assertion has failed. This could cause performance
1404 // problems for a method after its first do_null_assert failure.
1405 // Consider using 'Reason_class_check' instead?
1406
1407 // To cause an implicit null check, we set the not-null probability
1408 // to the maximum (PROB_MAX). For an explicit check the probability
1409 // is set to a smaller value.
1410 if (null_control != NULL || too_many_traps(reason)) {
1411 // probability is less likely
1412 ok_prob = PROB_LIKELY_MAG(3);
1413 } else if (!assert_null &&
1414 (ImplicitNullCheckThreshold > 0) &&
1415 method() != NULL &&
1416 (method()->method_data()->trap_count(reason)
1450 }
1451
1452 if (assert_null) {
1453 // Cast obj to null on this path.
1454 replace_in_map(value, zerocon(type));
1455 return zerocon(type);
1456 }
1457
1458 // Cast obj to not-null on this path, if there is no null_control.
1459 // (If there is a null_control, a non-null value may come back to haunt us.)
1460 if (type == T_OBJECT) {
1461 Node* cast = cast_not_null(value, false);
1462 if (null_control == NULL || (*null_control) == top())
1463 replace_in_map(value, cast);
1464 value = cast;
1465 }
1466
1467 return value;
1468 }
1469
1470 //------------------------------cast_not_null----------------------------------
1471 // Cast obj to not-null on this path
1472 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) {
1473 if (obj->is_InlineType()) {
1474 InlineTypeNode* vt = obj->clone()->as_InlineType();
1475 vt->set_is_init(_gvn);
1476 vt = _gvn.transform(vt)->as_InlineType();
1477 if (do_replace_in_map) {
1478 replace_in_map(obj, vt);
1479 }
1480 return vt;
1481 } else if (obj->is_InlineTypePtr()) {
1482 // Cast oop input instead
1483 Node* cast = cast_not_null(obj->as_InlineTypePtr()->get_oop(), do_replace_in_map);
1484 if (cast->is_top()) {
1485 // Always null
1486 return top();
1487 }
1488 // Create a new node with the casted oop input and is_init set
1489 InlineTypeBaseNode* vt = obj->clone()->as_InlineTypePtr();
1490 vt->set_oop(cast);
1491 vt->set_is_init(_gvn);
1492 vt = _gvn.transform(vt)->as_InlineTypePtr();
1493 if (do_replace_in_map) {
1494 replace_in_map(obj, vt);
1495 }
1496 return vt;
1497 }
1498 const Type *t = _gvn.type(obj);
1499 const Type *t_not_null = t->join_speculative(TypePtr::NOTNULL);
1500 // Object is already not-null?
1501 if( t == t_not_null ) return obj;
1502
1503 Node *cast = new CastPPNode(obj,t_not_null);
1504 cast->init_req(0, control());
1505 cast = _gvn.transform( cast );
1506
1507 // Scan for instances of 'obj' in the current JVM mapping.
1508 // These instances are known to be not-null after the test.
1509 if (do_replace_in_map)
1510 replace_in_map(obj, cast);
1511
1512 return cast; // Return casted value
1513 }
1514
1515 // Sometimes in intrinsics, we implicitly know an object is not null
1516 // (there's no actual null check) so we can cast it to not null. In
1517 // the course of optimizations, the input to the cast can become null.
1604 // These are layered on top of the factory methods in LoadNode and StoreNode,
1605 // and integrate with the parser's memory state and _gvn engine.
1606 //
1607
1608 // factory methods in "int adr_idx"
1609 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt,
1610 int adr_idx,
1611 MemNode::MemOrd mo,
1612 LoadNode::ControlDependency control_dependency,
1613 bool require_atomic_access,
1614 bool unaligned,
1615 bool mismatched,
1616 bool unsafe,
1617 uint8_t barrier_data) {
1618 assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" );
1619 const TypePtr* adr_type = NULL; // debug-mode-only argument
1620 debug_only(adr_type = C->get_adr_type(adr_idx));
1621 Node* mem = memory(adr_idx);
1622 Node* ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt, mo, control_dependency, require_atomic_access, unaligned, mismatched, unsafe, barrier_data);
1623 ld = _gvn.transform(ld);
1624
1625 if (((bt == T_OBJECT || bt == T_PRIMITIVE_OBJECT) && C->do_escape_analysis()) || C->eliminate_boxing()) {
1626 // Improve graph before escape analysis and boxing elimination.
1627 record_for_igvn(ld);
1628 }
1629 return ld;
1630 }
1631
1632 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt,
1633 int adr_idx,
1634 MemNode::MemOrd mo,
1635 bool require_atomic_access,
1636 bool unaligned,
1637 bool mismatched,
1638 bool unsafe) {
1639 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
1640 const TypePtr* adr_type = NULL;
1641 debug_only(adr_type = C->get_adr_type(adr_idx));
1642 Node *mem = memory(adr_idx);
1643 Node* st = StoreNode::make(_gvn, ctl, mem, adr, adr_type, val, bt, mo, require_atomic_access);
1644 if (unaligned) {
1645 st->as_Store()->set_unaligned_access();
1649 }
1650 if (unsafe) {
1651 st->as_Store()->set_unsafe_access();
1652 }
1653 st = _gvn.transform(st);
1654 set_memory(st, adr_idx);
1655 // Back-to-back stores can only remove intermediate store with DU info
1656 // so push on worklist for optimizer.
1657 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address))
1658 record_for_igvn(st);
1659
1660 return st;
1661 }
1662
1663 Node* GraphKit::access_store_at(Node* obj,
1664 Node* adr,
1665 const TypePtr* adr_type,
1666 Node* val,
1667 const Type* val_type,
1668 BasicType bt,
1669 DecoratorSet decorators,
1670 bool safe_for_replace) {
1671 // Transformation of a value which could be NULL pointer (CastPP #NULL)
1672 // could be delayed during Parse (for example, in adjust_map_after_if()).
1673 // Execute transformation here to avoid barrier generation in such case.
1674 if (_gvn.type(val) == TypePtr::NULL_PTR) {
1675 val = _gvn.makecon(TypePtr::NULL_PTR);
1676 }
1677
1678 if (stopped()) {
1679 return top(); // Dead path ?
1680 }
1681
1682 assert(val != NULL, "not dead path");
1683 if (val->is_InlineType()) {
1684 // Store to non-flattened field. Buffer the inline type and make sure
1685 // the store is re-executed if the allocation triggers deoptimization.
1686 PreserveReexecuteState preexecs(this);
1687 jvms()->set_should_reexecute(true);
1688 val = val->as_InlineType()->buffer(this, safe_for_replace);
1689 }
1690
1691 C2AccessValuePtr addr(adr, adr_type);
1692 C2AccessValue value(val, val_type);
1693 C2ParseAccess access(this, decorators | C2_WRITE_ACCESS, bt, obj, addr);
1694 if (access.is_raw()) {
1695 return _barrier_set->BarrierSetC2::store_at(access, value);
1696 } else {
1697 return _barrier_set->store_at(access, value);
1698 }
1699 }
1700
1701 Node* GraphKit::access_load_at(Node* obj, // containing obj
1702 Node* adr, // actual address to store val at
1703 const TypePtr* adr_type,
1704 const Type* val_type,
1705 BasicType bt,
1706 DecoratorSet decorators,
1707 Node* ctl) {
1708 if (stopped()) {
1709 return top(); // Dead path ?
1710 }
1711
1712 C2AccessValuePtr addr(adr, adr_type);
1713 C2ParseAccess access(this, decorators | C2_READ_ACCESS, bt, obj, addr, ctl);
1714 if (access.is_raw()) {
1715 return _barrier_set->BarrierSetC2::load_at(access, val_type);
1716 } else {
1717 return _barrier_set->load_at(access, val_type);
1718 }
1719 }
1720
1721 Node* GraphKit::access_load(Node* adr, // actual address to load val at
1722 const Type* val_type,
1723 BasicType bt,
1724 DecoratorSet decorators) {
1725 if (stopped()) {
1726 return top(); // Dead path ?
1727 }
1728
1729 C2AccessValuePtr addr(adr, adr->bottom_type()->is_ptr());
1730 C2ParseAccess access(this, decorators | C2_READ_ACCESS, bt, NULL, addr);
1731 if (access.is_raw()) {
1732 return _barrier_set->BarrierSetC2::load_at(access, val_type);
1733 } else {
1798 Node* new_val,
1799 const Type* value_type,
1800 BasicType bt,
1801 DecoratorSet decorators) {
1802 C2AccessValuePtr addr(adr, adr_type);
1803 C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS, bt, obj, addr, alias_idx);
1804 if (access.is_raw()) {
1805 return _barrier_set->BarrierSetC2::atomic_add_at(access, new_val, value_type);
1806 } else {
1807 return _barrier_set->atomic_add_at(access, new_val, value_type);
1808 }
1809 }
1810
1811 void GraphKit::access_clone(Node* src, Node* dst, Node* size, bool is_array) {
1812 return _barrier_set->clone(this, src, dst, size, is_array);
1813 }
1814
1815 //-------------------------array_element_address-------------------------
1816 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt,
1817 const TypeInt* sizetype, Node* ctrl) {
1818 const TypeAryPtr* arytype = _gvn.type(ary)->is_aryptr();
1819 uint shift = arytype->is_flat() ? arytype->flat_log_elem_size() : exact_log2(type2aelembytes(elembt));
1820 uint header = arrayOopDesc::base_offset_in_bytes(elembt);
1821
1822 // short-circuit a common case (saves lots of confusing waste motion)
1823 jint idx_con = find_int_con(idx, -1);
1824 if (idx_con >= 0) {
1825 intptr_t offset = header + ((intptr_t)idx_con << shift);
1826 return basic_plus_adr(ary, offset);
1827 }
1828
1829 // must be correct type for alignment purposes
1830 Node* base = basic_plus_adr(ary, header);
1831 idx = Compile::conv_I2X_index(&_gvn, idx, sizetype, ctrl);
1832 Node* scale = _gvn.transform( new LShiftXNode(idx, intcon(shift)) );
1833 return basic_plus_adr(ary, base, scale);
1834 }
1835
1836 //-------------------------load_array_element-------------------------
1837 Node* GraphKit::load_array_element(Node* ary, Node* idx, const TypeAryPtr* arytype, bool set_ctrl) {
1838 const Type* elemtype = arytype->elem();
1839 BasicType elembt = elemtype->array_element_basic_type();
1840 assert(elembt != T_PRIMITIVE_OBJECT, "inline types are not supported by this method");
1841 Node* adr = array_element_address(ary, idx, elembt, arytype->size());
1842 if (elembt == T_NARROWOOP) {
1843 elembt = T_OBJECT; // To satisfy switch in LoadNode::make()
1844 }
1845 Node* ld = access_load_at(ary, adr, arytype, elemtype, elembt,
1846 IN_HEAP | IS_ARRAY | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0));
1847 return ld;
1848 }
1849
1850 //-------------------------set_arguments_for_java_call-------------------------
1851 // Arguments (pre-popped from the stack) are taken from the JVMS.
1852 void GraphKit::set_arguments_for_java_call(CallJavaNode* call, bool is_late_inline) {
1853 PreserveReexecuteState preexecs(this);
1854 if (EnableValhalla) {
1855 // Make sure the call is "re-executed", if buffering of inline type arguments triggers deoptimization.
1856 // At this point, the call hasn't been executed yet, so we will only ever execute the call once.
1857 jvms()->set_should_reexecute(true);
1858 int arg_size = method()->get_declared_signature_at_bci(bci())->arg_size_for_bc(java_bc());
1859 inc_sp(arg_size);
1860 }
1861 // Add the call arguments
1862 const TypeTuple* domain = call->tf()->domain_sig();
1863 uint nargs = domain->cnt();
1864 int arg_num = 0;
1865 for (uint i = TypeFunc::Parms, idx = TypeFunc::Parms; i < nargs; i++) {
1866 Node* arg = argument(i-TypeFunc::Parms);
1867 const Type* t = domain->field_at(i);
1868 if (t->is_inlinetypeptr() && call->method()->is_scalarized_arg(arg_num)) {
1869 // We don't pass inline type arguments by reference but instead pass each field of the inline type
1870 if (!arg->is_InlineTypeBase()) {
1871 assert(_gvn.type(arg)->is_zero_type() && !t->inline_klass()->is_null_free(), "Unexpected argument type");
1872 arg = InlineTypeNode::make_from_oop(this, arg, t->inline_klass(), t->inline_klass()->is_null_free());
1873 }
1874 InlineTypeBaseNode* vt = arg->as_InlineTypeBase();
1875 vt->pass_fields(this, call, idx, true, !t->maybe_null());
1876 // If an inline type argument is passed as fields, attach the Method* to the call site
1877 // to be able to access the extended signature later via attached_method_before_pc().
1878 // For example, see CompiledMethod::preserve_callee_argument_oops().
1879 call->set_override_symbolic_info(true);
1880 arg_num++;
1881 continue;
1882 } else if (arg->is_InlineType()) {
1883 // Pass inline type argument via oop to callee
1884 arg = arg->as_InlineType()->buffer(this);
1885 if (!is_late_inline) {
1886 arg = arg->as_InlineTypePtr()->get_oop();
1887 }
1888 }
1889 if (t != Type::HALF) {
1890 arg_num++;
1891 }
1892 call->init_req(idx++, arg);
1893 }
1894 }
1895
1896 //---------------------------set_edges_for_java_call---------------------------
1897 // Connect a newly created call into the current JVMS.
1898 // A return value node (if any) is returned from set_edges_for_java_call.
1899 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) {
1900
1901 // Add the predefined inputs:
1902 call->init_req( TypeFunc::Control, control() );
1903 call->init_req( TypeFunc::I_O , i_o() );
1904 call->init_req( TypeFunc::Memory , reset_memory() );
1905 call->init_req( TypeFunc::FramePtr, frameptr() );
1906 call->init_req( TypeFunc::ReturnAdr, top() );
1907
1908 add_safepoint_edges(call, must_throw);
1909
1910 Node* xcall = _gvn.transform(call);
1911
1912 if (xcall == top()) {
1913 set_control(top());
1914 return;
1915 }
1916 assert(xcall == call, "call identity is stable");
1917
1918 // Re-use the current map to produce the result.
1919
1920 set_control(_gvn.transform(new ProjNode(call, TypeFunc::Control)));
1921 set_i_o( _gvn.transform(new ProjNode(call, TypeFunc::I_O , separate_io_proj)));
1922 set_all_memory_call(xcall, separate_io_proj);
1923
1924 //return xcall; // no need, caller already has it
1925 }
1926
1927 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj, bool deoptimize) {
1928 if (stopped()) return top(); // maybe the call folded up?
1929
1930 // Note: Since any out-of-line call can produce an exception,
1931 // we always insert an I_O projection from the call into the result.
1932
1933 make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj, deoptimize);
1934
1935 if (separate_io_proj) {
1936 // The caller requested separate projections be used by the fall
1937 // through and exceptional paths, so replace the projections for
1938 // the fall through path.
1939 set_i_o(_gvn.transform( new ProjNode(call, TypeFunc::I_O) ));
1940 set_all_memory(_gvn.transform( new ProjNode(call, TypeFunc::Memory) ));
1941 }
1942
1943 // Capture the return value, if any.
1944 Node* ret;
1945 if (call->method() == NULL || call->method()->return_type()->basic_type() == T_VOID) {
1946 ret = top();
1947 } else if (call->tf()->returns_inline_type_as_fields()) {
1948 // Return of multiple values (inline type fields): we create a
1949 // InlineType node, each field is a projection from the call.
1950 ciInlineKlass* vk = call->method()->return_type()->as_inline_klass();
1951 uint base_input = TypeFunc::Parms;
1952 ret = InlineTypeNode::make_from_multi(this, call, vk, base_input, false, call->method()->signature()->returns_null_free_inline_type());
1953 } else {
1954 ret = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
1955 }
1956
1957 return ret;
1958 }
1959
1960 //--------------------set_predefined_input_for_runtime_call--------------------
1961 // Reading and setting the memory state is way conservative here.
1962 // The real problem is that I am not doing real Type analysis on memory,
1963 // so I cannot distinguish card mark stores from other stores. Across a GC
1964 // point the Store Barrier and the card mark memory has to agree. I cannot
1965 // have a card mark store and its barrier split across the GC point from
1966 // either above or below. Here I get that to happen by reading ALL of memory.
1967 // A better answer would be to separate out card marks from other memory.
1968 // For now, return the input memory state, so that it can be reused
1969 // after the call, if this call has restricted memory effects.
1970 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call, Node* narrow_mem) {
1971 // Set fixed predefined input arguments
1972 Node* memory = reset_memory();
1973 Node* m = narrow_mem == NULL ? memory : narrow_mem;
1974 call->init_req( TypeFunc::Control, control() );
1975 call->init_req( TypeFunc::I_O, top() ); // does no i/o
1976 call->init_req( TypeFunc::Memory, m ); // may gc ptrs
2027 if (use->is_MergeMem()) {
2028 wl.push(use);
2029 }
2030 }
2031 }
2032
2033 // Replace the call with the current state of the kit.
2034 void GraphKit::replace_call(CallNode* call, Node* result, bool do_replaced_nodes) {
2035 JVMState* ejvms = NULL;
2036 if (has_exceptions()) {
2037 ejvms = transfer_exceptions_into_jvms();
2038 }
2039
2040 ReplacedNodes replaced_nodes = map()->replaced_nodes();
2041 ReplacedNodes replaced_nodes_exception;
2042 Node* ex_ctl = top();
2043
2044 SafePointNode* final_state = stop();
2045
2046 // Find all the needed outputs of this call
2047 CallProjections* callprojs = call->extract_projections(true);
2048
2049 Unique_Node_List wl;
2050 Node* init_mem = call->in(TypeFunc::Memory);
2051 Node* final_mem = final_state->in(TypeFunc::Memory);
2052 Node* final_ctl = final_state->in(TypeFunc::Control);
2053 Node* final_io = final_state->in(TypeFunc::I_O);
2054
2055 // Replace all the old call edges with the edges from the inlining result
2056 if (callprojs->fallthrough_catchproj != NULL) {
2057 C->gvn_replace_by(callprojs->fallthrough_catchproj, final_ctl);
2058 }
2059 if (callprojs->fallthrough_memproj != NULL) {
2060 if (final_mem->is_MergeMem()) {
2061 // Parser's exits MergeMem was not transformed but may be optimized
2062 final_mem = _gvn.transform(final_mem);
2063 }
2064 C->gvn_replace_by(callprojs->fallthrough_memproj, final_mem);
2065 add_mergemem_users_to_worklist(wl, final_mem);
2066 }
2067 if (callprojs->fallthrough_ioproj != NULL) {
2068 C->gvn_replace_by(callprojs->fallthrough_ioproj, final_io);
2069 }
2070
2071 // Replace the result with the new result if it exists and is used
2072 if (callprojs->resproj[0] != NULL && result != NULL) {
2073 // If the inlined code is dead, the result projections for an inline type returned as
2074 // fields have not been replaced. They will go away once the call is replaced by TOP below.
2075 assert(callprojs->nb_resproj == 1 || (call->tf()->returns_inline_type_as_fields() && stopped()),
2076 "unexpected number of results");
2077 C->gvn_replace_by(callprojs->resproj[0], result);
2078 }
2079
2080 if (ejvms == NULL) {
2081 // No exception edges to simply kill off those paths
2082 if (callprojs->catchall_catchproj != NULL) {
2083 C->gvn_replace_by(callprojs->catchall_catchproj, C->top());
2084 }
2085 if (callprojs->catchall_memproj != NULL) {
2086 C->gvn_replace_by(callprojs->catchall_memproj, C->top());
2087 }
2088 if (callprojs->catchall_ioproj != NULL) {
2089 C->gvn_replace_by(callprojs->catchall_ioproj, C->top());
2090 }
2091 // Replace the old exception object with top
2092 if (callprojs->exobj != NULL) {
2093 C->gvn_replace_by(callprojs->exobj, C->top());
2094 }
2095 } else {
2096 GraphKit ekit(ejvms);
2097
2098 // Load my combined exception state into the kit, with all phis transformed:
2099 SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states();
2100 replaced_nodes_exception = ex_map->replaced_nodes();
2101
2102 Node* ex_oop = ekit.use_exception_state(ex_map);
2103
2104 if (callprojs->catchall_catchproj != NULL) {
2105 C->gvn_replace_by(callprojs->catchall_catchproj, ekit.control());
2106 ex_ctl = ekit.control();
2107 }
2108 if (callprojs->catchall_memproj != NULL) {
2109 Node* ex_mem = ekit.reset_memory();
2110 C->gvn_replace_by(callprojs->catchall_memproj, ex_mem);
2111 add_mergemem_users_to_worklist(wl, ex_mem);
2112 }
2113 if (callprojs->catchall_ioproj != NULL) {
2114 C->gvn_replace_by(callprojs->catchall_ioproj, ekit.i_o());
2115 }
2116
2117 // Replace the old exception object with the newly created one
2118 if (callprojs->exobj != NULL) {
2119 C->gvn_replace_by(callprojs->exobj, ex_oop);
2120 }
2121 }
2122
2123 // Disconnect the call from the graph
2124 call->disconnect_inputs(C);
2125 C->gvn_replace_by(call, C->top());
2126
2127 // Clean up any MergeMems that feed other MergeMems since the
2128 // optimizer doesn't like that.
2129 while (wl.size() > 0) {
2130 _gvn.transform(wl.pop());
2131 }
2132
2133 if (callprojs->fallthrough_catchproj != NULL && !final_ctl->is_top() && do_replaced_nodes) {
2134 replaced_nodes.apply(C, final_ctl);
2135 }
2136 if (!ex_ctl->is_top() && do_replaced_nodes) {
2137 replaced_nodes_exception.apply(C, ex_ctl);
2138 }
2139 }
2140
2141
2142 //------------------------------increment_counter------------------------------
2143 // for statistics: increment a VM counter by 1
2144
2145 void GraphKit::increment_counter(address counter_addr) {
2146 Node* adr1 = makecon(TypeRawPtr::make(counter_addr));
2147 increment_counter(adr1);
2148 }
2149
2150 void GraphKit::increment_counter(Node* counter_addr) {
2151 int adr_type = Compile::AliasIdxRaw;
2152 Node* ctrl = control();
2153 Node* cnt = make_load(ctrl, counter_addr, TypeLong::LONG, T_LONG, adr_type, MemNode::unordered);
2312 *
2313 * @param n node that the type applies to
2314 * @param exact_kls type from profiling
2315 * @param maybe_null did profiling see null?
2316 *
2317 * @return node with improved type
2318 */
2319 Node* GraphKit::record_profile_for_speculation(Node* n, ciKlass* exact_kls, ProfilePtrKind ptr_kind) {
2320 const Type* current_type = _gvn.type(n);
2321 assert(UseTypeSpeculation, "type speculation must be on");
2322
2323 const TypePtr* speculative = current_type->speculative();
2324
2325 // Should the klass from the profile be recorded in the speculative type?
2326 if (current_type->would_improve_type(exact_kls, jvms()->depth())) {
2327 const TypeKlassPtr* tklass = TypeKlassPtr::make(exact_kls);
2328 const TypeOopPtr* xtype = tklass->as_instance_type();
2329 assert(xtype->klass_is_exact(), "Should be exact");
2330 // Any reason to believe n is not null (from this profiling or a previous one)?
2331 assert(ptr_kind != ProfileAlwaysNull, "impossible here");
2332 const TypePtr* ptr = (ptr_kind != ProfileNeverNull && current_type->speculative_maybe_null()) ? TypePtr::BOTTOM : TypePtr::NOTNULL;
2333 // record the new speculative type's depth
2334 speculative = xtype->cast_to_ptr_type(ptr->ptr())->is_ptr();
2335 speculative = speculative->with_inline_depth(jvms()->depth());
2336 } else if (current_type->would_improve_ptr(ptr_kind)) {
2337 // Profiling report that null was never seen so we can change the
2338 // speculative type to non null ptr.
2339 if (ptr_kind == ProfileAlwaysNull) {
2340 speculative = TypePtr::NULL_PTR;
2341 } else {
2342 assert(ptr_kind == ProfileNeverNull, "nothing else is an improvement");
2343 const TypePtr* ptr = TypePtr::NOTNULL;
2344 if (speculative != NULL) {
2345 speculative = speculative->cast_to_ptr_type(ptr->ptr())->is_ptr();
2346 } else {
2347 speculative = ptr;
2348 }
2349 }
2350 }
2351
2352 if (speculative != current_type->speculative()) {
2353 // Build a type with a speculative type (what we think we know
2354 // about the type but will need a guard when we use it)
2355 const TypeOopPtr* spec_type = TypeOopPtr::make(TypePtr::BotPTR, Type::Offset::bottom, TypeOopPtr::InstanceBot, speculative);
2356 // We're changing the type, we need a new CheckCast node to carry
2357 // the new type. The new type depends on the control: what
2358 // profiling tells us is only valid from here as far as we can
2359 // tell.
2360 Node* cast = new CheckCastPPNode(control(), n, current_type->remove_speculative()->join_speculative(spec_type));
2361 cast = _gvn.transform(cast);
2362 replace_in_map(n, cast);
2363 n = cast;
2364 }
2365
2366 return n;
2367 }
2368
2369 /**
2370 * Record profiling data from receiver profiling at an invoke with the
2371 * type system so that it can propagate it (speculation)
2372 *
2373 * @param n receiver node
2374 *
2375 * @return node with improved type
2376 */
2377 Node* GraphKit::record_profiled_receiver_for_speculation(Node* n) {
2378 if (!UseTypeSpeculation) {
2379 return n;
2380 }
2381 ciKlass* exact_kls = profile_has_unique_klass();
2382 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2383 if ((java_bc() == Bytecodes::_checkcast ||
2384 java_bc() == Bytecodes::_instanceof ||
2385 java_bc() == Bytecodes::_aastore) &&
2386 method()->method_data()->is_mature()) {
2387 ciProfileData* data = method()->method_data()->bci_to_data(bci());
2388 if (data != NULL) {
2389 if (java_bc() == Bytecodes::_aastore) {
2390 ciKlass* array_type = NULL;
2391 ciKlass* element_type = NULL;
2392 ProfilePtrKind element_ptr = ProfileMaybeNull;
2393 bool flat_array = true;
2394 bool null_free_array = true;
2395 method()->array_access_profiled_type(bci(), array_type, element_type, element_ptr, flat_array, null_free_array);
2396 exact_kls = element_type;
2397 ptr_kind = element_ptr;
2398 } else {
2399 if (!data->as_BitData()->null_seen()) {
2400 ptr_kind = ProfileNeverNull;
2401 } else {
2402 assert(data->is_ReceiverTypeData(), "bad profile data type");
2403 ciReceiverTypeData* call = (ciReceiverTypeData*)data->as_ReceiverTypeData();
2404 uint i = 0;
2405 for (; i < call->row_limit(); i++) {
2406 ciKlass* receiver = call->receiver(i);
2407 if (receiver != NULL) {
2408 break;
2409 }
2410 }
2411 ptr_kind = (i == call->row_limit()) ? ProfileAlwaysNull : ProfileMaybeNull;
2412 }
2413 }
2414 }
2415 }
2416 return record_profile_for_speculation(n, exact_kls, ptr_kind);
2417 }
2418
2419 /**
2420 * Record profiling data from argument profiling at an invoke with the
2421 * type system so that it can propagate it (speculation)
2422 *
2423 * @param dest_method target method for the call
2424 * @param bc what invoke bytecode is this?
2425 */
2426 void GraphKit::record_profiled_arguments_for_speculation(ciMethod* dest_method, Bytecodes::Code bc) {
2427 if (!UseTypeSpeculation) {
2428 return;
2429 }
2430 const TypeFunc* tf = TypeFunc::make(dest_method);
2431 int nargs = tf->domain_sig()->cnt() - TypeFunc::Parms;
2432 int skip = Bytecodes::has_receiver(bc) ? 1 : 0;
2433 for (int j = skip, i = 0; j < nargs && i < TypeProfileArgsLimit; j++) {
2434 const Type *targ = tf->domain_sig()->field_at(j + TypeFunc::Parms);
2435 if (is_reference_type(targ->basic_type())) {
2436 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2437 ciKlass* better_type = NULL;
2438 if (method()->argument_profiled_type(bci(), i, better_type, ptr_kind)) {
2439 record_profile_for_speculation(argument(j), better_type, ptr_kind);
2440 }
2441 i++;
2442 }
2443 }
2444 }
2445
2446 /**
2447 * Record profiling data from parameter profiling at an invoke with
2448 * the type system so that it can propagate it (speculation)
2449 */
2450 void GraphKit::record_profiled_parameters_for_speculation() {
2451 if (!UseTypeSpeculation) {
2452 return;
2453 }
2454 for (int i = 0, j = 0; i < method()->arg_size() ; i++) {
2468 * the type system so that it can propagate it (speculation)
2469 */
2470 void GraphKit::record_profiled_return_for_speculation() {
2471 if (!UseTypeSpeculation) {
2472 return;
2473 }
2474 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2475 ciKlass* better_type = NULL;
2476 if (method()->return_profiled_type(bci(), better_type, ptr_kind)) {
2477 // If profiling reports a single type for the return value,
2478 // feed it to the type system so it can propagate it as a
2479 // speculative type
2480 record_profile_for_speculation(stack(sp()-1), better_type, ptr_kind);
2481 }
2482 }
2483
2484 void GraphKit::round_double_arguments(ciMethod* dest_method) {
2485 if (Matcher::strict_fp_requires_explicit_rounding) {
2486 // (Note: TypeFunc::make has a cache that makes this fast.)
2487 const TypeFunc* tf = TypeFunc::make(dest_method);
2488 int nargs = tf->domain_sig()->cnt() - TypeFunc::Parms;
2489 for (int j = 0; j < nargs; j++) {
2490 const Type *targ = tf->domain_sig()->field_at(j + TypeFunc::Parms);
2491 if (targ->basic_type() == T_DOUBLE) {
2492 // If any parameters are doubles, they must be rounded before
2493 // the call, dprecision_rounding does gvn.transform
2494 Node *arg = argument(j);
2495 arg = dprecision_rounding(arg);
2496 set_argument(j, arg);
2497 }
2498 }
2499 }
2500 }
2501
2502 // rounding for strict float precision conformance
2503 Node* GraphKit::precision_rounding(Node* n) {
2504 if (Matcher::strict_fp_requires_explicit_rounding) {
2505 #ifdef IA32
2506 if (UseSSE == 0) {
2507 return _gvn.transform(new RoundFloatNode(0, n));
2508 }
2509 #else
2510 Unimplemented();
2619 // The first NULL ends the list.
2620 Node* parm0, Node* parm1,
2621 Node* parm2, Node* parm3,
2622 Node* parm4, Node* parm5,
2623 Node* parm6, Node* parm7) {
2624 assert(call_addr != NULL, "must not call NULL targets");
2625
2626 // Slow-path call
2627 bool is_leaf = !(flags & RC_NO_LEAF);
2628 bool has_io = (!is_leaf && !(flags & RC_NO_IO));
2629 if (call_name == NULL) {
2630 assert(!is_leaf, "must supply name for leaf");
2631 call_name = OptoRuntime::stub_name(call_addr);
2632 }
2633 CallNode* call;
2634 if (!is_leaf) {
2635 call = new CallStaticJavaNode(call_type, call_addr, call_name, adr_type);
2636 } else if (flags & RC_NO_FP) {
2637 call = new CallLeafNoFPNode(call_type, call_addr, call_name, adr_type);
2638 } else if (flags & RC_VECTOR){
2639 uint num_bits = call_type->range_sig()->field_at(TypeFunc::Parms)->is_vect()->length_in_bytes() * BitsPerByte;
2640 call = new CallLeafVectorNode(call_type, call_addr, call_name, adr_type, num_bits);
2641 } else {
2642 call = new CallLeafNode(call_type, call_addr, call_name, adr_type);
2643 }
2644
2645 // The following is similar to set_edges_for_java_call,
2646 // except that the memory effects of the call are restricted to AliasIdxRaw.
2647
2648 // Slow path call has no side-effects, uses few values
2649 bool wide_in = !(flags & RC_NARROW_MEM);
2650 bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot);
2651
2652 Node* prev_mem = NULL;
2653 if (wide_in) {
2654 prev_mem = set_predefined_input_for_runtime_call(call);
2655 } else {
2656 assert(!wide_out, "narrow in => narrow out");
2657 Node* narrow_mem = memory(adr_type);
2658 prev_mem = set_predefined_input_for_runtime_call(call, narrow_mem);
2659 }
2699
2700 if (has_io) {
2701 set_i_o(_gvn.transform(new ProjNode(call, TypeFunc::I_O)));
2702 }
2703 return call;
2704
2705 }
2706
2707 // i2b
2708 Node* GraphKit::sign_extend_byte(Node* in) {
2709 Node* tmp = _gvn.transform(new LShiftINode(in, _gvn.intcon(24)));
2710 return _gvn.transform(new RShiftINode(tmp, _gvn.intcon(24)));
2711 }
2712
2713 // i2s
2714 Node* GraphKit::sign_extend_short(Node* in) {
2715 Node* tmp = _gvn.transform(new LShiftINode(in, _gvn.intcon(16)));
2716 return _gvn.transform(new RShiftINode(tmp, _gvn.intcon(16)));
2717 }
2718
2719
2720 //------------------------------merge_memory-----------------------------------
2721 // Merge memory from one path into the current memory state.
2722 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) {
2723 for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) {
2724 Node* old_slice = mms.force_memory();
2725 Node* new_slice = mms.memory2();
2726 if (old_slice != new_slice) {
2727 PhiNode* phi;
2728 if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region) {
2729 if (mms.is_empty()) {
2730 // clone base memory Phi's inputs for this memory slice
2731 assert(old_slice == mms.base_memory(), "sanity");
2732 phi = PhiNode::make(region, NULL, Type::MEMORY, mms.adr_type(C));
2733 _gvn.set_type(phi, Type::MEMORY);
2734 for (uint i = 1; i < phi->req(); i++) {
2735 phi->init_req(i, old_slice->in(i));
2736 }
2737 } else {
2738 phi = old_slice->as_Phi(); // Phi was generated already
2739 }
2951
2952 // Now do a linear scan of the secondary super-klass array. Again, no real
2953 // performance impact (too rare) but it's gotta be done.
2954 // Since the code is rarely used, there is no penalty for moving it
2955 // out of line, and it can only improve I-cache density.
2956 // The decision to inline or out-of-line this final check is platform
2957 // dependent, and is found in the AD file definition of PartialSubtypeCheck.
2958 Node* psc = gvn.transform(
2959 new PartialSubtypeCheckNode(*ctrl, subklass, superklass));
2960
2961 IfNode *iff4 = gen_subtype_check_compare(*ctrl, psc, gvn.zerocon(T_OBJECT), BoolTest::ne, PROB_FAIR, gvn, T_ADDRESS);
2962 r_not_subtype->init_req(2, gvn.transform(new IfTrueNode (iff4)));
2963 r_ok_subtype ->init_req(3, gvn.transform(new IfFalseNode(iff4)));
2964
2965 // Return false path; set default control to true path.
2966 *ctrl = gvn.transform(r_ok_subtype);
2967 return gvn.transform(r_not_subtype);
2968 }
2969
2970 Node* GraphKit::gen_subtype_check(Node* obj_or_subklass, Node* superklass) {
2971 const Type* sub_t = _gvn.type(obj_or_subklass);
2972 if (sub_t->isa_inlinetype()) {
2973 obj_or_subklass = makecon(TypeKlassPtr::make(sub_t->inline_klass()));
2974 }
2975 bool expand_subtype_check = C->post_loop_opts_phase() || // macro node expansion is over
2976 ExpandSubTypeCheckAtParseTime; // forced expansion
2977 if (expand_subtype_check) {
2978 MergeMemNode* mem = merged_memory();
2979 Node* ctrl = control();
2980 Node* subklass = obj_or_subklass;
2981 if (!sub_t->isa_klassptr() && !sub_t->isa_inlinetype()) {
2982 subklass = load_object_klass(obj_or_subklass);
2983 }
2984 Node* n = Phase::gen_subtype_check(subklass, superklass, &ctrl, mem, _gvn);
2985 set_control(ctrl);
2986 return n;
2987 }
2988
2989 Node* check = _gvn.transform(new SubTypeCheckNode(C, obj_or_subklass, superklass));
2990 Node* bol = _gvn.transform(new BoolNode(check, BoolTest::eq));
2991 IfNode* iff = create_and_xform_if(control(), bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
2992 set_control(_gvn.transform(new IfTrueNode(iff)));
2993 return _gvn.transform(new IfFalseNode(iff));
2994 }
2995
2996 // Profile-driven exact type check:
2997 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass,
2998 float prob, Node* *casted_receiver) {
2999 assert(!klass->is_interface(), "no exact type check on interfaces");
3000 Node* fail = top();
3001 const Type* rec_t = _gvn.type(receiver);
3002 if (rec_t->isa_inlinetype()) {
3003 if (klass->equals(rec_t->inline_klass())) {
3004 (*casted_receiver) = receiver; // Always passes
3005 } else {
3006 (*casted_receiver) = top(); // Always fails
3007 fail = control();
3008 set_control(top());
3009 }
3010 return fail;
3011 }
3012 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass);
3013 Node* recv_klass = load_object_klass(receiver);
3014 fail = type_check(recv_klass, tklass, prob);
3015
3016 if (!stopped()) {
3017 const TypeOopPtr* receiver_type = _gvn.type(receiver)->isa_oopptr();
3018 const TypeOopPtr* recv_xtype = tklass->as_instance_type();
3019 assert(recv_xtype->klass_is_exact(), "");
3020
3021 if (!receiver_type->higher_equal(recv_xtype)) { // ignore redundant casts
3022 // Subsume downstream occurrences of receiver with a cast to
3023 // recv_xtype, since now we know what the type will be.
3024 Node* cast = new CheckCastPPNode(control(), receiver, recv_xtype);
3025 Node* res = _gvn.transform(cast);
3026 if (recv_xtype->is_inlinetypeptr()) {
3027 assert(!gvn().type(res)->maybe_null(), "receiver should never be null");
3028 res = InlineTypeNode::make_from_oop(this, res, recv_xtype->inline_klass())->as_InlineTypeBase()->as_ptr(&gvn());
3029 }
3030 (*casted_receiver) = res;
3031 // (User must make the replace_in_map call.)
3032 }
3033 }
3034
3035 return fail;
3036 }
3037
3038 Node* GraphKit::type_check(Node* recv_klass, const TypeKlassPtr* tklass,
3039 float prob) {
3040 Node* want_klass = makecon(tklass);
3041 Node* cmp = _gvn.transform(new CmpPNode(recv_klass, want_klass));
3042 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
3043 IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN);
3044 set_control(_gvn.transform(new IfTrueNode (iff)));
3045 Node* fail = _gvn.transform(new IfFalseNode(iff));
3046 return fail;
3047 }
3048
3049 //------------------------------subtype_check_receiver-------------------------
3050 Node* GraphKit::subtype_check_receiver(Node* receiver, ciKlass* klass,
3051 Node** casted_receiver) {
3052 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass);
3053 Node* want_klass = makecon(tklass);
3054
3055 Node* slow_ctl = gen_subtype_check(receiver, want_klass);
3056
3057 // Ignore interface type information until interface types are properly tracked.
3058 if (!stopped() && !klass->is_interface()) {
3059 const TypeOopPtr* receiver_type = _gvn.type(receiver)->isa_oopptr();
3060 const TypeOopPtr* recv_type = tklass->cast_to_exactness(false)->is_klassptr()->as_instance_type();
3061 if (receiver_type != NULL && !receiver_type->higher_equal(recv_type)) { // ignore redundant casts
3062 Node* cast = new CheckCastPPNode(control(), receiver, recv_type);
3063 (*casted_receiver) = _gvn.transform(cast);
3064 }
3065 }
3066
3067 return slow_ctl;
3068 }
3069
3070 //------------------------------seems_never_null-------------------------------
3071 // Use null_seen information if it is available from the profile.
3072 // If we see an unexpected null at a type check we record it and force a
3073 // recompile; the offending check will be recompiled to handle NULLs.
3074 // If we see several offending BCIs, then all checks in the
3075 // method will be recompiled.
3076 bool GraphKit::seems_never_null(Node* obj, ciProfileData* data, bool& speculating) {
3077 speculating = !_gvn.type(obj)->speculative_maybe_null();
3078 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculating);
3079 if (UncommonNullCast // Cutout for this technique
3080 && obj != null() // And not the -Xcomp stupid case?
3081 && !too_many_traps(reason)
3082 ) {
3083 if (speculating) {
3084 return true;
3085 }
3086 if (data == NULL)
3087 // Edge case: no mature data. Be optimistic here.
3088 return true;
3089 // If the profile has not seen a null, assume it won't happen.
3090 assert(java_bc() == Bytecodes::_checkcast ||
3091 java_bc() == Bytecodes::_instanceof ||
3092 java_bc() == Bytecodes::_aastore, "MDO must collect null_seen bit here");
3093 if (java_bc() == Bytecodes::_aastore) {
3094 return ((ciArrayLoadStoreData*)data->as_ArrayLoadStoreData())->element()->ptr_kind() == ProfileNeverNull;
3095 }
3096 return !data->as_BitData()->null_seen();
3097 }
3098 speculating = false;
3099 return false;
3100 }
3101
3102 void GraphKit::guard_klass_being_initialized(Node* klass) {
3103 int init_state_off = in_bytes(InstanceKlass::init_state_offset());
3104 Node* adr = basic_plus_adr(top(), klass, init_state_off);
3105 Node* init_state = LoadNode::make(_gvn, NULL, immutable_memory(), adr,
3106 adr->bottom_type()->is_ptr(), TypeInt::BYTE,
3107 T_BYTE, MemNode::unordered);
3108 init_state = _gvn.transform(init_state);
3109
3110 Node* being_initialized_state = makecon(TypeInt::make(InstanceKlass::being_initialized));
3111
3112 Node* chk = _gvn.transform(new CmpINode(being_initialized_state, init_state));
3113 Node* tst = _gvn.transform(new BoolNode(chk, BoolTest::eq));
3114
3115 { BuildCutout unless(this, tst, PROB_MAX);
3155
3156 //------------------------maybe_cast_profiled_receiver-------------------------
3157 // If the profile has seen exactly one type, narrow to exactly that type.
3158 // Subsequent type checks will always fold up.
3159 Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj,
3160 const TypeKlassPtr* require_klass,
3161 ciKlass* spec_klass,
3162 bool safe_for_replace) {
3163 if (!UseTypeProfile || !TypeProfileCasts) return NULL;
3164
3165 Deoptimization::DeoptReason reason = Deoptimization::reason_class_check(spec_klass != NULL);
3166
3167 // Make sure we haven't already deoptimized from this tactic.
3168 if (too_many_traps_or_recompiles(reason))
3169 return NULL;
3170
3171 // (No, this isn't a call, but it's enough like a virtual call
3172 // to use the same ciMethod accessor to get the profile info...)
3173 // If we have a speculative type use it instead of profiling (which
3174 // may not help us)
3175 ciKlass* exact_kls = spec_klass;
3176 if (exact_kls == NULL) {
3177 if (java_bc() == Bytecodes::_aastore) {
3178 ciKlass* array_type = NULL;
3179 ciKlass* element_type = NULL;
3180 ProfilePtrKind element_ptr = ProfileMaybeNull;
3181 bool flat_array = true;
3182 bool null_free_array = true;
3183 method()->array_access_profiled_type(bci(), array_type, element_type, element_ptr, flat_array, null_free_array);
3184 exact_kls = element_type;
3185 } else {
3186 exact_kls = profile_has_unique_klass();
3187 }
3188 }
3189 if (exact_kls != NULL) {// no cast failures here
3190 if (require_klass == NULL ||
3191 C->static_subtype_check(require_klass, TypeKlassPtr::make(exact_kls)) == Compile::SSC_always_true) {
3192 // If we narrow the type to match what the type profile sees or
3193 // the speculative type, we can then remove the rest of the
3194 // cast.
3195 // This is a win, even if the exact_kls is very specific,
3196 // because downstream operations, such as method calls,
3197 // will often benefit from the sharper type.
3198 Node* exact_obj = not_null_obj; // will get updated in place...
3199 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0,
3200 &exact_obj);
3201 { PreserveJVMState pjvms(this);
3202 set_control(slow_ctl);
3203 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile);
3204 }
3205 if (safe_for_replace) {
3206 replace_in_map(not_null_obj, exact_obj);
3207 }
3208 return exact_obj;
3273 // and the reflective instance-of call.
3274 Node* GraphKit::gen_instanceof(Node* obj, Node* superklass, bool safe_for_replace) {
3275 kill_dead_locals(); // Benefit all the uncommon traps
3276 assert( !stopped(), "dead parse path should be checked in callers" );
3277 assert(!TypePtr::NULL_PTR->higher_equal(_gvn.type(superklass)->is_klassptr()),
3278 "must check for not-null not-dead klass in callers");
3279
3280 // Make the merge point
3281 enum { _obj_path = 1, _fail_path, _null_path, PATH_LIMIT };
3282 RegionNode* region = new RegionNode(PATH_LIMIT);
3283 Node* phi = new PhiNode(region, TypeInt::BOOL);
3284 C->set_has_split_ifs(true); // Has chance for split-if optimization
3285
3286 ciProfileData* data = NULL;
3287 if (java_bc() == Bytecodes::_instanceof) { // Only for the bytecode
3288 data = method()->method_data()->bci_to_data(bci());
3289 }
3290 bool speculative_not_null = false;
3291 bool never_see_null = (ProfileDynamicTypes // aggressive use of profile
3292 && seems_never_null(obj, data, speculative_not_null));
3293 bool is_value = obj->is_InlineType();
3294
3295 // Null check; get casted pointer; set region slot 3
3296 Node* null_ctl = top();
3297 if (is_value) {
3298 // TODO 8284443 Enable this
3299 safe_for_replace = false;
3300 never_see_null = false;
3301 }
3302 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null);
3303
3304 // If not_null_obj is dead, only null-path is taken
3305 if (stopped()) { // Doing instance-of on a NULL?
3306 set_control(null_ctl);
3307 return intcon(0);
3308 }
3309 region->init_req(_null_path, null_ctl);
3310 phi ->init_req(_null_path, intcon(0)); // Set null path value
3311 if (null_ctl == top()) {
3312 // Do this eagerly, so that pattern matches like is_diamond_phi
3313 // will work even during parsing.
3314 assert(_null_path == PATH_LIMIT-1, "delete last");
3315 region->del_req(_null_path);
3316 phi ->del_req(_null_path);
3317 }
3318
3319 // Do we know the type check always succeed?
3320 if (!is_value) {
3321 bool known_statically = false;
3322 if (_gvn.type(superklass)->singleton()) {
3323 const TypeKlassPtr* superk = _gvn.type(superklass)->is_klassptr();
3324 const TypeKlassPtr* subk = _gvn.type(obj)->is_oopptr()->as_klass_type();
3325 if (subk != NULL && subk->is_loaded()) {
3326 int static_res = C->static_subtype_check(superk, subk);
3327 known_statically = (static_res == Compile::SSC_always_true || static_res == Compile::SSC_always_false);
3328 }
3329 }
3330
3331 if (!known_statically) {
3332 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3333 // We may not have profiling here or it may not help us. If we
3334 // have a speculative type use it to perform an exact cast.
3335 ciKlass* spec_obj_type = obj_type->speculative_type();
3336 if (spec_obj_type != NULL || (ProfileDynamicTypes && data != NULL)) {
3337 Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, NULL, spec_obj_type, safe_for_replace);
3338 if (stopped()) { // Profile disagrees with this path.
3339 set_control(null_ctl); // Null is the only remaining possibility.
3340 return intcon(0);
3341 }
3342 if (cast_obj != NULL) {
3343 not_null_obj = cast_obj;
3344 is_value = not_null_obj->is_InlineType();
3345 }
3346 }
3347 }
3348 }
3349
3350 // Generate the subtype check
3351 Node* not_subtype_ctrl = gen_subtype_check(not_null_obj, superklass);
3352
3353 // Plug in the success path to the general merge in slot 1.
3354 region->init_req(_obj_path, control());
3355 phi ->init_req(_obj_path, intcon(1));
3356
3357 // Plug in the failing path to the general merge in slot 2.
3358 region->init_req(_fail_path, not_subtype_ctrl);
3359 phi ->init_req(_fail_path, intcon(0));
3360
3361 // Return final merged results
3362 set_control( _gvn.transform(region) );
3363 record_for_igvn(region);
3364
3365 // If we know the type check always succeeds then we don't use the
3366 // profiling data at this bytecode. Don't lose it, feed it to the
3367 // type system as a speculative type.
3368 if (safe_for_replace && !is_value) {
3369 Node* casted_obj = record_profiled_receiver_for_speculation(obj);
3370 replace_in_map(obj, casted_obj);
3371 }
3372
3373 return _gvn.transform(phi);
3374 }
3375
3376 //-------------------------------gen_checkcast---------------------------------
3377 // Generate a checkcast idiom. Used by both the checkcast bytecode and the
3378 // array store bytecode. Stack must be as-if BEFORE doing the bytecode so the
3379 // uncommon-trap paths work. Adjust stack after this call.
3380 // If failure_control is supplied and not null, it is filled in with
3381 // the control edge for the cast failure. Otherwise, an appropriate
3382 // uncommon trap or exception is thrown.
3383 Node* GraphKit::gen_checkcast(Node *obj, Node* superklass, Node* *failure_control, bool null_free) {
3384 kill_dead_locals(); // Benefit all the uncommon traps
3385 const TypeKlassPtr* tk = _gvn.type(superklass)->is_klassptr();
3386 const TypeOopPtr* toop = tk->cast_to_exactness(false)->as_instance_type();
3387 bool safe_for_replace = (failure_control == NULL);
3388 bool from_inline = obj->is_InlineType();
3389 assert(!null_free || toop->is_inlinetypeptr(), "must be an inline type pointer");
3390
3391 // Fast cutout: Check the case that the cast is vacuously true.
3392 // This detects the common cases where the test will short-circuit
3393 // away completely. We do this before we perform the null check,
3394 // because if the test is going to turn into zero code, we don't
3395 // want a residual null check left around. (Causes a slowdown,
3396 // for example, in some objArray manipulations, such as a[i]=a[j].)
3397 if (tk->singleton()) {
3398 const TypeKlassPtr* kptr = NULL;
3399 const Type* t = _gvn.type(obj);
3400 if (t->isa_oop_ptr()) {
3401 kptr = t->is_oopptr()->as_klass_type();
3402 } else if (obj->is_InlineTypeBase()) {
3403 ciInlineKlass* vk = t->inline_klass();
3404 kptr = TypeInstKlassPtr::make(TypePtr::NotNull, vk, Type::Offset(0), vk->flatten_array());
3405 }
3406 if (kptr != NULL) {
3407 switch (C->static_subtype_check(tk, kptr)) {
3408 case Compile::SSC_always_true:
3409 // If we know the type check always succeed then we don't use
3410 // the profiling data at this bytecode. Don't lose it, feed it
3411 // to the type system as a speculative type.
3412 if (!from_inline) {
3413 obj = record_profiled_receiver_for_speculation(obj);
3414 }
3415 if (null_free) {
3416 assert(safe_for_replace, "must be");
3417 obj = null_check(obj);
3418 }
3419 assert(stopped() || !toop->is_inlinetypeptr() || obj->is_InlineTypeBase(), "should have been scalarized");
3420 return obj;
3421 case Compile::SSC_always_false:
3422 if (null_free) {
3423 assert(safe_for_replace, "must be");
3424 obj = null_check(obj);
3425 }
3426 // It needs a null check because a null will *pass* the cast check.
3427 if (t->isa_oopptr() != NULL && !t->is_oopptr()->maybe_null()) {
3428 bool is_aastore = (java_bc() == Bytecodes::_aastore);
3429 Deoptimization::DeoptReason reason = is_aastore ?
3430 Deoptimization::Reason_array_check : Deoptimization::Reason_class_check;
3431 builtin_throw(reason);
3432 return top();
3433 } else if (!too_many_traps_or_recompiles(Deoptimization::Reason_null_assert)) {
3434 return null_assert(obj);
3435 }
3436 break; // Fall through to full check
3437 default:
3438 break;
3439 }
3440 }
3441 }
3442
3443 ciProfileData* data = NULL;
3444 if (failure_control == NULL) { // use MDO in regular case only
3445 assert(java_bc() == Bytecodes::_aastore ||
3446 java_bc() == Bytecodes::_checkcast,
3447 "interpreter profiles type checks only for these BCs");
3448 if (method()->method_data()->is_mature()) {
3449 data = method()->method_data()->bci_to_data(bci());
3450 }
3451 }
3452
3453 // Make the merge point
3454 enum { _obj_path = 1, _null_path, PATH_LIMIT };
3455 RegionNode* region = new RegionNode(PATH_LIMIT);
3456 Node* phi = new PhiNode(region, toop);
3457 _gvn.set_type(region, Type::CONTROL);
3458 _gvn.set_type(phi, toop);
3459
3460 C->set_has_split_ifs(true); // Has chance for split-if optimization
3461
3462 // Use null-cast information if it is available
3463 bool speculative_not_null = false;
3464 bool never_see_null = ((failure_control == NULL) // regular case only
3465 && seems_never_null(obj, data, speculative_not_null));
3466
3467 // Null check; get casted pointer; set region slot 3
3468 Node* null_ctl = top();
3469 Node* not_null_obj = NULL;
3470 if (null_free) {
3471 assert(safe_for_replace, "must be");
3472 not_null_obj = null_check(obj);
3473 } else if (from_inline) {
3474 // TODO 8284443 obj can be null and null should pass
3475 not_null_obj = obj;
3476 } else {
3477 not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null);
3478 }
3479
3480 // If not_null_obj is dead, only null-path is taken
3481 if (stopped()) { // Doing instance-of on a NULL?
3482 set_control(null_ctl);
3483 if (toop->is_inlinetypeptr()) {
3484 return InlineTypePtrNode::make_null(_gvn, toop->inline_klass());
3485 }
3486 return null();
3487 }
3488 region->init_req(_null_path, null_ctl);
3489 phi ->init_req(_null_path, null()); // Set null path value
3490 if (null_ctl == top()) {
3491 // Do this eagerly, so that pattern matches like is_diamond_phi
3492 // will work even during parsing.
3493 assert(_null_path == PATH_LIMIT-1, "delete last");
3494 region->del_req(_null_path);
3495 phi ->del_req(_null_path);
3496 }
3497
3498 Node* cast_obj = NULL;
3499 if (!from_inline && tk->klass_is_exact()) {
3500 // The following optimization tries to statically cast the speculative type of the object
3501 // (for example obtained during profiling) to the type of the superklass and then do a
3502 // dynamic check that the type of the object is what we expect. To work correctly
3503 // for checkcast and aastore the type of superklass should be exact.
3504 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3505 // We may not have profiling here or it may not help us. If we have
3506 // a speculative type use it to perform an exact cast.
3507 ciKlass* spec_obj_type = obj_type->speculative_type();
3508 if (spec_obj_type != NULL || data != NULL) {
3509 cast_obj = maybe_cast_profiled_receiver(not_null_obj, tk, spec_obj_type, safe_for_replace);
3510 if (cast_obj != NULL) {
3511 if (failure_control != NULL) // failure is now impossible
3512 (*failure_control) = top();
3513 // adjust the type of the phi to the exact klass:
3514 phi->raise_bottom_type(_gvn.type(cast_obj)->meet_speculative(TypePtr::NULL_PTR));
3515 }
3516 }
3517 }
3518
3519 if (cast_obj == NULL) {
3520 // Generate the subtype check
3521 Node* not_subtype_ctrl = gen_subtype_check(not_null_obj, superklass);
3522
3523 // Plug in success path into the merge
3524 cast_obj = from_inline ? not_null_obj : _gvn.transform(new CheckCastPPNode(control(), not_null_obj, toop));
3525 // Failure path ends in uncommon trap (or may be dead - failure impossible)
3526 if (failure_control == NULL) {
3527 if (not_subtype_ctrl != top()) { // If failure is possible
3528 PreserveJVMState pjvms(this);
3529 set_control(not_subtype_ctrl);
3530 Node* obj_klass = NULL;
3531 if (not_null_obj->is_InlineTypeBase()) {
3532 obj_klass = makecon(TypeKlassPtr::make(_gvn.type(not_null_obj)->inline_klass()));
3533 } else {
3534 obj_klass = load_object_klass(not_null_obj);
3535 }
3536 bool is_aastore = (java_bc() == Bytecodes::_aastore);
3537 Deoptimization::DeoptReason reason = is_aastore ?
3538 Deoptimization::Reason_array_check : Deoptimization::Reason_class_check;
3539 builtin_throw(reason);
3540 }
3541 } else {
3542 (*failure_control) = not_subtype_ctrl;
3543 }
3544 }
3545
3546 region->init_req(_obj_path, control());
3547 phi ->init_req(_obj_path, cast_obj);
3548
3549 // A merge of NULL or Casted-NotNull obj
3550 Node* res = _gvn.transform(phi);
3551
3552 // Note I do NOT always 'replace_in_map(obj,result)' here.
3553 // if( tk->klass()->can_be_primary_super() )
3554 // This means that if I successfully store an Object into an array-of-String
3555 // I 'forget' that the Object is really now known to be a String. I have to
3556 // do this because we don't have true union types for interfaces - if I store
3557 // a Baz into an array-of-Interface and then tell the optimizer it's an
3558 // Interface, I forget that it's also a Baz and cannot do Baz-like field
3559 // references to it. FIX THIS WHEN UNION TYPES APPEAR!
3560 // replace_in_map( obj, res );
3561
3562 // Return final merged results
3563 set_control( _gvn.transform(region) );
3564 record_for_igvn(region);
3565
3566 bool not_inline = !toop->can_be_inline_type();
3567 bool not_flattened = !UseFlatArray || not_inline || (toop->is_inlinetypeptr() && !toop->inline_klass()->flatten_array());
3568 if (EnableValhalla && not_flattened) {
3569 // Check if obj has been loaded from an array
3570 obj = obj->isa_DecodeN() ? obj->in(1) : obj;
3571 Node* array = NULL;
3572 if (obj->isa_Load()) {
3573 Node* address = obj->in(MemNode::Address);
3574 if (address->isa_AddP()) {
3575 array = address->as_AddP()->in(AddPNode::Base);
3576 }
3577 } else if (obj->is_Phi()) {
3578 Node* region = obj->in(0);
3579 // TODO make this more robust (see JDK-8231346)
3580 if (region->req() == 3 && region->in(2) != NULL && region->in(2)->in(0) != NULL) {
3581 IfNode* iff = region->in(2)->in(0)->isa_If();
3582 if (iff != NULL) {
3583 iff->is_flat_array_check(&_gvn, &array);
3584 }
3585 }
3586 }
3587 if (array != NULL) {
3588 const TypeAryPtr* ary_t = _gvn.type(array)->isa_aryptr();
3589 if (ary_t != NULL) {
3590 if (!ary_t->is_not_null_free() && not_inline) {
3591 // Casting array element to a non-inline-type, mark array as not null-free.
3592 Node* cast = _gvn.transform(new CheckCastPPNode(control(), array, ary_t->cast_to_not_null_free()));
3593 replace_in_map(array, cast);
3594 } else if (!ary_t->is_not_flat()) {
3595 // Casting array element to a non-flattened type, mark array as not flat.
3596 Node* cast = _gvn.transform(new CheckCastPPNode(control(), array, ary_t->cast_to_not_flat()));
3597 replace_in_map(array, cast);
3598 }
3599 }
3600 }
3601 }
3602
3603 if (!stopped() && !res->is_InlineTypeBase()) {
3604 res = record_profiled_receiver_for_speculation(res);
3605 if (toop->is_inlinetypeptr()) {
3606 Node* vt = InlineTypeNode::make_from_oop(this, res, toop->inline_klass(), !gvn().type(res)->maybe_null());
3607 res = vt;
3608 if (safe_for_replace) {
3609 if (vt->is_InlineType() && C->inlining_incrementally()) {
3610 vt = vt->as_InlineType()->as_ptr(&_gvn);
3611 }
3612 replace_in_map(obj, vt);
3613 replace_in_map(not_null_obj, vt);
3614 replace_in_map(res, vt);
3615 }
3616 }
3617 }
3618 return res;
3619 }
3620
3621 Node* GraphKit::inline_type_test(Node* obj, bool is_inline) {
3622 Node* mark_adr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
3623 Node* mark = make_load(NULL, mark_adr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered);
3624 Node* mask = MakeConX(markWord::inline_type_pattern);
3625 Node* masked = _gvn.transform(new AndXNode(mark, mask));
3626 Node* cmp = _gvn.transform(new CmpXNode(masked, mask));
3627 return _gvn.transform(new BoolNode(cmp, is_inline ? BoolTest::eq : BoolTest::ne));
3628 }
3629
3630 Node* GraphKit::is_val_mirror(Node* mirror) {
3631 Node* p = basic_plus_adr(mirror, java_lang_Class::secondary_mirror_offset());
3632 Node* secondary_mirror = access_load_at(mirror, p, _gvn.type(p)->is_ptr(), TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR), T_OBJECT, IN_HEAP);
3633 Node* cmp = _gvn.transform(new CmpPNode(mirror, secondary_mirror));
3634 return _gvn.transform(new BoolNode(cmp, BoolTest::eq));
3635 }
3636
3637 Node* GraphKit::array_lh_test(Node* klass, jint mask, jint val, bool eq) {
3638 Node* lh_adr = basic_plus_adr(klass, in_bytes(Klass::layout_helper_offset()));
3639 // Make sure to use immutable memory here to enable hoisting the check out of loops
3640 Node* lh_val = _gvn.transform(LoadNode::make(_gvn, NULL, immutable_memory(), lh_adr, lh_adr->bottom_type()->is_ptr(), TypeInt::INT, T_INT, MemNode::unordered));
3641 Node* masked = _gvn.transform(new AndINode(lh_val, intcon(mask)));
3642 Node* cmp = _gvn.transform(new CmpINode(masked, intcon(val)));
3643 return _gvn.transform(new BoolNode(cmp, eq ? BoolTest::eq : BoolTest::ne));
3644 }
3645
3646 Node* GraphKit::flat_array_test(Node* array_or_klass, bool flat) {
3647 // We can't use immutable memory here because the mark word is mutable.
3648 // PhaseIdealLoop::move_flat_array_check_out_of_loop will make sure the
3649 // check is moved out of loops (mainly to enable loop unswitching).
3650 Node* mem = UseArrayMarkWordCheck ? memory(Compile::AliasIdxRaw) : immutable_memory();
3651 Node* cmp = _gvn.transform(new FlatArrayCheckNode(C, mem, array_or_klass));
3652 record_for_igvn(cmp); // Give it a chance to be optimized out by IGVN
3653 return _gvn.transform(new BoolNode(cmp, flat ? BoolTest::eq : BoolTest::ne));
3654 }
3655
3656 Node* GraphKit::null_free_array_test(Node* klass, bool null_free) {
3657 return array_lh_test(klass, Klass::_lh_null_free_array_bit_inplace, 0, !null_free);
3658 }
3659
3660 // Deoptimize if 'ary' is a null-free inline type array and 'val' is null
3661 Node* GraphKit::inline_array_null_guard(Node* ary, Node* val, int nargs, bool safe_for_replace) {
3662 RegionNode* region = new RegionNode(3);
3663 Node* null_ctl = top();
3664 null_check_oop(val, &null_ctl);
3665 if (null_ctl != top()) {
3666 PreserveJVMState pjvms(this);
3667 set_control(null_ctl);
3668 {
3669 // Deoptimize if null-free array
3670 BuildCutout unless(this, null_free_array_test(load_object_klass(ary), /* null_free = */ false), PROB_MAX);
3671 inc_sp(nargs);
3672 uncommon_trap(Deoptimization::Reason_null_check,
3673 Deoptimization::Action_none);
3674 }
3675 region->init_req(1, control());
3676 }
3677 region->init_req(2, control());
3678 set_control(_gvn.transform(region));
3679 record_for_igvn(region);
3680 if (_gvn.type(val) == TypePtr::NULL_PTR) {
3681 // Since we were just successfully storing null, the array can't be null free.
3682 const TypeAryPtr* ary_t = _gvn.type(ary)->is_aryptr();
3683 ary_t = ary_t->cast_to_not_null_free();
3684 Node* cast = _gvn.transform(new CheckCastPPNode(control(), ary, ary_t));
3685 if (safe_for_replace) {
3686 replace_in_map(ary, cast);
3687 }
3688 ary = cast;
3689 }
3690 return ary;
3691 }
3692
3693 //------------------------------next_monitor-----------------------------------
3694 // What number should be given to the next monitor?
3695 int GraphKit::next_monitor() {
3696 int current = jvms()->monitor_depth()* C->sync_stack_slots();
3697 int next = current + C->sync_stack_slots();
3698 // Keep the toplevel high water mark current:
3699 if (C->fixed_slots() < next) C->set_fixed_slots(next);
3700 return current;
3701 }
3702
3703 //------------------------------insert_mem_bar---------------------------------
3704 // Memory barrier to avoid floating things around
3705 // The membar serves as a pinch point between both control and all memory slices.
3706 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) {
3707 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent);
3708 mb->init_req(TypeFunc::Control, control());
3709 mb->init_req(TypeFunc::Memory, reset_memory());
3710 Node* membar = _gvn.transform(mb);
3738 }
3739 Node* membar = _gvn.transform(mb);
3740 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control)));
3741 if (alias_idx == Compile::AliasIdxBot) {
3742 merged_memory()->set_base_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)));
3743 } else {
3744 set_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)),alias_idx);
3745 }
3746 return membar;
3747 }
3748
3749 //------------------------------shared_lock------------------------------------
3750 // Emit locking code.
3751 FastLockNode* GraphKit::shared_lock(Node* obj) {
3752 // bci is either a monitorenter bc or InvocationEntryBci
3753 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
3754 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
3755
3756 if( !GenerateSynchronizationCode )
3757 return NULL; // Not locking things?
3758
3759 if (stopped()) // Dead monitor?
3760 return NULL;
3761
3762 assert(dead_locals_are_killed(), "should kill locals before sync. point");
3763
3764 // Box the stack location
3765 Node* box = _gvn.transform(new BoxLockNode(next_monitor()));
3766 Node* mem = reset_memory();
3767
3768 FastLockNode * flock = _gvn.transform(new FastLockNode(0, obj, box) )->as_FastLock();
3769
3770 // Create the rtm counters for this fast lock if needed.
3771 flock->create_rtm_lock_counter(sync_jvms()); // sync_jvms used to get current bci
3772
3773 // Add monitor to debug info for the slow path. If we block inside the
3774 // slow path and de-opt, we need the monitor hanging around
3775 map()->push_monitor( flock );
3776
3777 const TypeFunc *tf = LockNode::lock_type();
3778 LockNode *lock = new LockNode(C, tf);
3807 }
3808 #endif
3809
3810 return flock;
3811 }
3812
3813
3814 //------------------------------shared_unlock----------------------------------
3815 // Emit unlocking code.
3816 void GraphKit::shared_unlock(Node* box, Node* obj) {
3817 // bci is either a monitorenter bc or InvocationEntryBci
3818 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
3819 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
3820
3821 if( !GenerateSynchronizationCode )
3822 return;
3823 if (stopped()) { // Dead monitor?
3824 map()->pop_monitor(); // Kill monitor from debug info
3825 return;
3826 }
3827 assert(!obj->is_InlineTypeBase(), "should not unlock on inline type");
3828
3829 // Memory barrier to avoid floating things down past the locked region
3830 insert_mem_bar(Op_MemBarReleaseLock);
3831
3832 const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type();
3833 UnlockNode *unlock = new UnlockNode(C, tf);
3834 #ifdef ASSERT
3835 unlock->set_dbg_jvms(sync_jvms());
3836 #endif
3837 uint raw_idx = Compile::AliasIdxRaw;
3838 unlock->init_req( TypeFunc::Control, control() );
3839 unlock->init_req( TypeFunc::Memory , memory(raw_idx) );
3840 unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
3841 unlock->init_req( TypeFunc::FramePtr, frameptr() );
3842 unlock->init_req( TypeFunc::ReturnAdr, top() );
3843
3844 unlock->init_req(TypeFunc::Parms + 0, obj);
3845 unlock->init_req(TypeFunc::Parms + 1, box);
3846 unlock = _gvn.transform(unlock)->as_Unlock();
3847
3848 Node* mem = reset_memory();
3849
3850 // unlock has no side-effects, sets few values
3851 set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM);
3852
3853 // Kill monitor from debug info
3854 map()->pop_monitor( );
3855 }
3856
3857 //-------------------------------get_layout_helper-----------------------------
3858 // If the given klass is a constant or known to be an array,
3859 // fetch the constant layout helper value into constant_value
3860 // and return (Node*)NULL. Otherwise, load the non-constant
3861 // layout helper value, and return the node which represents it.
3862 // This two-faced routine is useful because allocation sites
3863 // almost always feature constant types.
3864 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) {
3865 const TypeKlassPtr* inst_klass = _gvn.type(klass_node)->isa_klassptr();
3866 if (!StressReflectiveCode && inst_klass != NULL) {
3867 bool xklass = inst_klass->klass_is_exact();
3868 bool can_be_flattened = false;
3869 const TypeAryPtr* ary_type = inst_klass->as_instance_type()->isa_aryptr();
3870 if (UseFlatArray && !xklass && ary_type != NULL && !ary_type->is_null_free()) {
3871 // The runtime type of [LMyValue might be [QMyValue due to [QMyValue <: [LMyValue. Don't constant fold.
3872 const TypeOopPtr* elem = ary_type->elem()->make_oopptr();
3873 can_be_flattened = ary_type->can_be_inline_array() && (!elem->is_inlinetypeptr() || elem->inline_klass()->flatten_array());
3874 }
3875 if (!can_be_flattened && (xklass || inst_klass->isa_aryklassptr())) {
3876 jint lhelper;
3877 if (inst_klass->is_flat()) {
3878 lhelper = ary_type->flat_layout_helper();
3879 } else if (inst_klass->isa_aryklassptr()) {
3880 BasicType elem = ary_type->elem()->array_element_basic_type();
3881 if (is_reference_type(elem, true)) {
3882 elem = T_OBJECT;
3883 }
3884 lhelper = Klass::array_layout_helper(elem);
3885 } else {
3886 lhelper = inst_klass->is_instklassptr()->exact_klass()->layout_helper();
3887 }
3888 if (lhelper != Klass::_lh_neutral_value) {
3889 constant_value = lhelper;
3890 return (Node*) NULL;
3891 }
3892 }
3893 }
3894 constant_value = Klass::_lh_neutral_value; // put in a known value
3895 Node* lhp = basic_plus_adr(klass_node, klass_node, in_bytes(Klass::layout_helper_offset()));
3896 return make_load(NULL, lhp, TypeInt::INT, T_INT, MemNode::unordered);
3897 }
3898
3899 // We just put in an allocate/initialize with a big raw-memory effect.
3900 // Hook selected additional alias categories on the initialization.
3901 static void hook_memory_on_init(GraphKit& kit, int alias_idx,
3902 MergeMemNode* init_in_merge,
3903 Node* init_out_raw) {
3904 DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory());
3905 assert(init_in_merge->memory_at(alias_idx) == init_in_raw, "");
3906
3907 Node* prevmem = kit.memory(alias_idx);
3908 init_in_merge->set_memory_at(alias_idx, prevmem);
3909 if (init_out_raw != NULL) {
3910 kit.set_memory(init_out_raw, alias_idx);
3911 }
3912 }
3913
3914 //---------------------------set_output_for_allocation-------------------------
3915 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc,
3916 const TypeOopPtr* oop_type,
3917 bool deoptimize_on_exception) {
3918 int rawidx = Compile::AliasIdxRaw;
3919 alloc->set_req( TypeFunc::FramePtr, frameptr() );
3920 add_safepoint_edges(alloc);
3921 Node* allocx = _gvn.transform(alloc);
3922 set_control( _gvn.transform(new ProjNode(allocx, TypeFunc::Control) ) );
3923 // create memory projection for i_o
3924 set_memory ( _gvn.transform( new ProjNode(allocx, TypeFunc::Memory, true) ), rawidx );
3925 make_slow_call_ex(allocx, env()->Throwable_klass(), true, deoptimize_on_exception);
3926
3927 // create a memory projection as for the normal control path
3928 Node* malloc = _gvn.transform(new ProjNode(allocx, TypeFunc::Memory));
3929 set_memory(malloc, rawidx);
3930
3931 // a normal slow-call doesn't change i_o, but an allocation does
3932 // we create a separate i_o projection for the normal control path
3933 set_i_o(_gvn.transform( new ProjNode(allocx, TypeFunc::I_O, false) ) );
3934 Node* rawoop = _gvn.transform( new ProjNode(allocx, TypeFunc::Parms) );
3935
3936 // put in an initialization barrier
3937 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx,
3938 rawoop)->as_Initialize();
3939 assert(alloc->initialization() == init, "2-way macro link must work");
3940 assert(init ->allocation() == alloc, "2-way macro link must work");
3941 {
3942 // Extract memory strands which may participate in the new object's
3943 // initialization, and source them from the new InitializeNode.
3944 // This will allow us to observe initializations when they occur,
3945 // and link them properly (as a group) to the InitializeNode.
3946 assert(init->in(InitializeNode::Memory) == malloc, "");
3947 MergeMemNode* minit_in = MergeMemNode::make(malloc);
3948 init->set_req(InitializeNode::Memory, minit_in);
3949 record_for_igvn(minit_in); // fold it up later, if possible
3950 _gvn.set_type(minit_in, Type::MEMORY);
3951 Node* minit_out = memory(rawidx);
3952 assert(minit_out->is_Proj() && minit_out->in(0) == init, "");
3953 // Add an edge in the MergeMem for the header fields so an access
3954 // to one of those has correct memory state
3955 set_memory(minit_out, C->get_alias_index(oop_type->add_offset(oopDesc::mark_offset_in_bytes())));
3956 set_memory(minit_out, C->get_alias_index(oop_type->add_offset(oopDesc::klass_offset_in_bytes())));
3957 if (oop_type->isa_aryptr()) {
3958 const TypeAryPtr* arytype = oop_type->is_aryptr();
3959 if (arytype->is_flat()) {
3960 // Initially all flattened array accesses share a single slice
3961 // but that changes after parsing. Prepare the memory graph so
3962 // it can optimize flattened array accesses properly once they
3963 // don't share a single slice.
3964 assert(C->flattened_accesses_share_alias(), "should be set at parse time");
3965 C->set_flattened_accesses_share_alias(false);
3966 ciInlineKlass* vk = arytype->elem()->inline_klass();
3967 for (int i = 0, len = vk->nof_nonstatic_fields(); i < len; i++) {
3968 ciField* field = vk->nonstatic_field_at(i);
3969 if (field->offset() >= TrackedInitializationLimit * HeapWordSize)
3970 continue; // do not bother to track really large numbers of fields
3971 int off_in_vt = field->offset() - vk->first_field_offset();
3972 const TypePtr* adr_type = arytype->with_field_offset(off_in_vt)->add_offset(Type::OffsetBot);
3973 int fieldidx = C->get_alias_index(adr_type, true);
3974 // Pass NULL for init_out. Having per flat array element field memory edges as uses of the Initialize node
3975 // can result in per flat array field Phis to be created which confuses the logic of
3976 // Compile::adjust_flattened_array_access_aliases().
3977 hook_memory_on_init(*this, fieldidx, minit_in, NULL);
3978 }
3979 C->set_flattened_accesses_share_alias(true);
3980 hook_memory_on_init(*this, C->get_alias_index(TypeAryPtr::INLINES), minit_in, minit_out);
3981 } else {
3982 const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot);
3983 int elemidx = C->get_alias_index(telemref);
3984 hook_memory_on_init(*this, elemidx, minit_in, minit_out);
3985 }
3986 } else if (oop_type->isa_instptr()) {
3987 set_memory(minit_out, C->get_alias_index(oop_type)); // mark word
3988 ciInstanceKlass* ik = oop_type->is_instptr()->instance_klass();
3989 for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) {
3990 ciField* field = ik->nonstatic_field_at(i);
3991 if (field->offset() >= TrackedInitializationLimit * HeapWordSize)
3992 continue; // do not bother to track really large numbers of fields
3993 // Find (or create) the alias category for this field:
3994 int fieldidx = C->alias_type(field)->index();
3995 hook_memory_on_init(*this, fieldidx, minit_in, minit_out);
3996 }
3997 }
3998 }
3999
4000 // Cast raw oop to the real thing...
4001 Node* javaoop = new CheckCastPPNode(control(), rawoop, oop_type);
4002 javaoop = _gvn.transform(javaoop);
4003 C->set_recent_alloc(control(), javaoop);
4004 assert(just_allocated_object(control()) == javaoop, "just allocated");
4005
4006 #ifdef ASSERT
4007 { // Verify that the AllocateNode::Ideal_allocation recognizers work:
4018 assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please");
4019 }
4020 }
4021 #endif //ASSERT
4022
4023 return javaoop;
4024 }
4025
4026 //---------------------------new_instance--------------------------------------
4027 // This routine takes a klass_node which may be constant (for a static type)
4028 // or may be non-constant (for reflective code). It will work equally well
4029 // for either, and the graph will fold nicely if the optimizer later reduces
4030 // the type to a constant.
4031 // The optional arguments are for specialized use by intrinsics:
4032 // - If 'extra_slow_test' if not null is an extra condition for the slow-path.
4033 // - If 'return_size_val', report the total object size to the caller.
4034 // - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize)
4035 Node* GraphKit::new_instance(Node* klass_node,
4036 Node* extra_slow_test,
4037 Node* *return_size_val,
4038 bool deoptimize_on_exception,
4039 InlineTypeBaseNode* inline_type_node) {
4040 // Compute size in doublewords
4041 // The size is always an integral number of doublewords, represented
4042 // as a positive bytewise size stored in the klass's layout_helper.
4043 // The layout_helper also encodes (in a low bit) the need for a slow path.
4044 jint layout_con = Klass::_lh_neutral_value;
4045 Node* layout_val = get_layout_helper(klass_node, layout_con);
4046 bool layout_is_con = (layout_val == NULL);
4047
4048 if (extra_slow_test == NULL) extra_slow_test = intcon(0);
4049 // Generate the initial go-slow test. It's either ALWAYS (return a
4050 // Node for 1) or NEVER (return a NULL) or perhaps (in the reflective
4051 // case) a computed value derived from the layout_helper.
4052 Node* initial_slow_test = NULL;
4053 if (layout_is_con) {
4054 assert(!StressReflectiveCode, "stress mode does not use these paths");
4055 bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con);
4056 initial_slow_test = must_go_slow ? intcon(1) : extra_slow_test;
4057 } else { // reflective case
4058 // This reflective path is used by Unsafe.allocateInstance.
4059 // (It may be stress-tested by specifying StressReflectiveCode.)
4060 // Basically, we want to get into the VM is there's an illegal argument.
4061 Node* bit = intcon(Klass::_lh_instance_slow_path_bit);
4062 initial_slow_test = _gvn.transform( new AndINode(layout_val, bit) );
4063 if (extra_slow_test != intcon(0)) {
4064 initial_slow_test = _gvn.transform( new OrINode(initial_slow_test, extra_slow_test) );
4065 }
4066 // (Macro-expander will further convert this to a Bool, if necessary.)
4077
4078 // Clear the low bits to extract layout_helper_size_in_bytes:
4079 assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit");
4080 Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong));
4081 size = _gvn.transform( new AndXNode(size, mask) );
4082 }
4083 if (return_size_val != NULL) {
4084 (*return_size_val) = size;
4085 }
4086
4087 // This is a precise notnull oop of the klass.
4088 // (Actually, it need not be precise if this is a reflective allocation.)
4089 // It's what we cast the result to.
4090 const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr();
4091 if (!tklass) tklass = TypeInstKlassPtr::OBJECT;
4092 const TypeOopPtr* oop_type = tklass->as_instance_type();
4093
4094 // Now generate allocation code
4095
4096 // The entire memory state is needed for slow path of the allocation
4097 // since GC and deoptimization can happen.
4098 Node *mem = reset_memory();
4099 set_all_memory(mem); // Create new memory state
4100
4101 AllocateNode* alloc = new AllocateNode(C, AllocateNode::alloc_type(Type::TOP),
4102 control(), mem, i_o(),
4103 size, klass_node,
4104 initial_slow_test, inline_type_node);
4105
4106 return set_output_for_allocation(alloc, oop_type, deoptimize_on_exception);
4107 }
4108
4109 //-------------------------------new_array-------------------------------------
4110 // helper for newarray and anewarray
4111 // The 'length' parameter is (obviously) the length of the array.
4112 // See comments on new_instance for the meaning of the other arguments.
4113 Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable)
4114 Node* length, // number of array elements
4115 int nargs, // number of arguments to push back for uncommon trap
4116 Node* *return_size_val,
4117 bool deoptimize_on_exception) {
4118 jint layout_con = Klass::_lh_neutral_value;
4119 Node* layout_val = get_layout_helper(klass_node, layout_con);
4120 bool layout_is_con = (layout_val == NULL);
4121
4122 if (!layout_is_con && !StressReflectiveCode &&
4123 !too_many_traps(Deoptimization::Reason_class_check)) {
4124 // This is a reflective array creation site.
4125 // Optimistically assume that it is a subtype of Object[],
4126 // so that we can fold up all the address arithmetic.
4127 layout_con = Klass::array_layout_helper(T_OBJECT);
4128 Node* cmp_lh = _gvn.transform( new CmpINode(layout_val, intcon(layout_con)) );
4129 Node* bol_lh = _gvn.transform( new BoolNode(cmp_lh, BoolTest::eq) );
4130 { BuildCutout unless(this, bol_lh, PROB_MAX);
4131 inc_sp(nargs);
4132 uncommon_trap(Deoptimization::Reason_class_check,
4133 Deoptimization::Action_maybe_recompile);
4134 }
4135 layout_val = NULL;
4136 layout_is_con = true;
4137 }
4138
4139 // Generate the initial go-slow test. Make sure we do not overflow
4140 // if length is huge (near 2Gig) or negative! We do not need
4141 // exact double-words here, just a close approximation of needed
4142 // double-words. We can't add any offset or rounding bits, lest we
4143 // take a size -1 of bytes and make it positive. Use an unsigned
4144 // compare, so negative sizes look hugely positive.
4145 int fast_size_limit = FastAllocateSizeLimit;
4146 if (layout_is_con) {
4147 assert(!StressReflectiveCode, "stress mode does not use these paths");
4148 // Increase the size limit if we have exact knowledge of array type.
4149 int log2_esize = Klass::layout_helper_log2_element_size(layout_con);
4150 fast_size_limit <<= MAX2(LogBytesPerLong - log2_esize, 0);
4151 }
4152
4153 Node* initial_slow_cmp = _gvn.transform( new CmpUNode( length, intcon( fast_size_limit ) ) );
4154 Node* initial_slow_test = _gvn.transform( new BoolNode( initial_slow_cmp, BoolTest::gt ) );
4155
4156 // --- Size Computation ---
4157 // array_size = round_to_heap(array_header + (length << elem_shift));
4158 // where round_to_heap(x) == align_to(x, MinObjAlignmentInBytes)
4159 // and align_to(x, y) == ((x + y-1) & ~(y-1))
4160 // The rounding mask is strength-reduced, if possible.
4161 int round_mask = MinObjAlignmentInBytes - 1;
4162 Node* header_size = NULL;
4163 int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
4164 // (T_BYTE has the weakest alignment and size restrictions...)
4165 if (layout_is_con) {
4166 int hsize = Klass::layout_helper_header_size(layout_con);
4167 int eshift = Klass::layout_helper_log2_element_size(layout_con);
4168 bool is_flat_array = Klass::layout_helper_is_flatArray(layout_con);
4169 if ((round_mask & ~right_n_bits(eshift)) == 0)
4170 round_mask = 0; // strength-reduce it if it goes away completely
4171 assert(is_flat_array || (hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded");
4172 assert(header_size_min <= hsize, "generic minimum is smallest");
4173 header_size_min = hsize;
4174 header_size = intcon(hsize + round_mask);
4175 } else {
4176 Node* hss = intcon(Klass::_lh_header_size_shift);
4177 Node* hsm = intcon(Klass::_lh_header_size_mask);
4178 Node* hsize = _gvn.transform( new URShiftINode(layout_val, hss) );
4179 hsize = _gvn.transform( new AndINode(hsize, hsm) );
4180 Node* mask = intcon(round_mask);
4181 header_size = _gvn.transform( new AddINode(hsize, mask) );
4182 }
4183
4184 Node* elem_shift = NULL;
4185 if (layout_is_con) {
4186 int eshift = Klass::layout_helper_log2_element_size(layout_con);
4187 if (eshift != 0)
4188 elem_shift = intcon(eshift);
4189 } else {
4190 // There is no need to mask or shift this value.
4191 // The semantics of LShiftINode include an implicit mask to 0x1F.
4235 // places, one where the length is sharply limited, and the other
4236 // after a successful allocation.
4237 Node* abody = lengthx;
4238 if (elem_shift != NULL)
4239 abody = _gvn.transform( new LShiftXNode(lengthx, elem_shift) );
4240 Node* size = _gvn.transform( new AddXNode(headerx, abody) );
4241 if (round_mask != 0) {
4242 Node* mask = MakeConX(~round_mask);
4243 size = _gvn.transform( new AndXNode(size, mask) );
4244 }
4245 // else if round_mask == 0, the size computation is self-rounding
4246
4247 if (return_size_val != NULL) {
4248 // This is the size
4249 (*return_size_val) = size;
4250 }
4251
4252 // Now generate allocation code
4253
4254 // The entire memory state is needed for slow path of the allocation
4255 // since GC and deoptimization can happen.
4256 Node *mem = reset_memory();
4257 set_all_memory(mem); // Create new memory state
4258
4259 if (initial_slow_test->is_Bool()) {
4260 // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick.
4261 initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn);
4262 }
4263
4264 const TypeKlassPtr* ary_klass = _gvn.type(klass_node)->isa_klassptr();
4265 const TypeOopPtr* ary_type = ary_klass->as_instance_type();
4266 const TypeAryPtr* ary_ptr = ary_type->isa_aryptr();
4267
4268 // Inline type array variants:
4269 // - null-ok: MyValue.ref[] (ciObjArrayKlass "[LMyValue")
4270 // - null-free: MyValue.val[] (ciObjArrayKlass "[QMyValue")
4271 // - null-free, flattened: MyValue.val[] (ciFlatArrayKlass "[QMyValue")
4272 // Check if array is a null-free, non-flattened inline type array
4273 // that needs to be initialized with the default inline type.
4274 Node* default_value = NULL;
4275 Node* raw_default_value = NULL;
4276 if (ary_ptr != NULL && ary_ptr->klass_is_exact()) {
4277 // Array type is known
4278 if (ary_ptr->is_null_free() && !ary_ptr->is_flat()) {
4279 ciInlineKlass* vk = ary_ptr->elem()->make_oopptr()->inline_klass();
4280 default_value = InlineTypeNode::default_oop(gvn(), vk);
4281 }
4282 } else if (ary_type->can_be_inline_array()) {
4283 // Array type is not known, add runtime checks
4284 assert(!ary_klass->klass_is_exact(), "unexpected exact type");
4285 Node* r = new RegionNode(3);
4286 default_value = new PhiNode(r, TypeInstPtr::BOTTOM);
4287
4288 Node* bol = array_lh_test(klass_node, Klass::_lh_array_tag_flat_value_bit_inplace | Klass::_lh_null_free_array_bit_inplace, Klass::_lh_null_free_array_bit_inplace);
4289 IfNode* iff = create_and_map_if(control(), bol, PROB_FAIR, COUNT_UNKNOWN);
4290
4291 // Null-free, non-flattened inline type array, initialize with the default value
4292 set_control(_gvn.transform(new IfTrueNode(iff)));
4293 Node* p = basic_plus_adr(klass_node, in_bytes(ArrayKlass::element_klass_offset()));
4294 Node* eklass = _gvn.transform(LoadKlassNode::make(_gvn, control(), immutable_memory(), p, TypeInstPtr::KLASS));
4295 Node* adr_fixed_block_addr = basic_plus_adr(eklass, in_bytes(InstanceKlass::adr_inlineklass_fixed_block_offset()));
4296 Node* adr_fixed_block = make_load(control(), adr_fixed_block_addr, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
4297 Node* default_value_offset_addr = basic_plus_adr(adr_fixed_block, in_bytes(InlineKlass::default_value_offset_offset()));
4298 Node* default_value_offset = make_load(control(), default_value_offset_addr, TypeInt::INT, T_INT, MemNode::unordered);
4299 Node* elem_mirror = load_mirror_from_klass(eklass);
4300 Node* default_value_addr = basic_plus_adr(elem_mirror, ConvI2X(default_value_offset));
4301 Node* val = access_load_at(elem_mirror, default_value_addr, TypeInstPtr::MIRROR, TypeInstPtr::NOTNULL, T_OBJECT, IN_HEAP);
4302 r->init_req(1, control());
4303 default_value->init_req(1, val);
4304
4305 // Otherwise initialize with all zero
4306 r->init_req(2, _gvn.transform(new IfFalseNode(iff)));
4307 default_value->init_req(2, null());
4308
4309 set_control(_gvn.transform(r));
4310 default_value = _gvn.transform(default_value);
4311 }
4312 if (default_value != NULL) {
4313 if (UseCompressedOops) {
4314 // With compressed oops, the 64-bit init value is built from two 32-bit compressed oops
4315 default_value = _gvn.transform(new EncodePNode(default_value, default_value->bottom_type()->make_narrowoop()));
4316 Node* lower = _gvn.transform(new CastP2XNode(control(), default_value));
4317 Node* upper = _gvn.transform(new LShiftLNode(lower, intcon(32)));
4318 raw_default_value = _gvn.transform(new OrLNode(lower, upper));
4319 } else {
4320 raw_default_value = _gvn.transform(new CastP2XNode(control(), default_value));
4321 }
4322 }
4323
4324 // Create the AllocateArrayNode and its result projections
4325 AllocateArrayNode* alloc
4326 = new AllocateArrayNode(C, AllocateArrayNode::alloc_type(TypeInt::INT),
4327 control(), mem, i_o(),
4328 size, klass_node,
4329 initial_slow_test,
4330 length,
4331 default_value, raw_default_value);
4332 // Cast to correct type. Note that the klass_node may be constant or not,
4333 // and in the latter case the actual array type will be inexact also.
4334 // (This happens via a non-constant argument to inline_native_newArray.)
4335 // In any case, the value of klass_node provides the desired array type.
4336 const TypeInt* length_type = _gvn.find_int_type(length);
4337 if (ary_type->isa_aryptr() && length_type != NULL) {
4338 // Try to get a better type than POS for the size
4339 ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
4340 }
4341
4342 Node* javaoop = set_output_for_allocation(alloc, ary_type, deoptimize_on_exception);
4343
4344 array_ideal_length(alloc, ary_type, true);
4345 return javaoop;
4346 }
4347
4348 // The following "Ideal_foo" functions are placed here because they recognize
4349 // the graph shapes created by the functions immediately above.
4350
4351 //---------------------------Ideal_allocation----------------------------------
4352 // Given an oop pointer or raw pointer, see if it feeds from an AllocateNode.
4353 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase) {
4354 if (ptr == NULL) { // reduce dumb test in callers
4355 return NULL;
4356 }
4465 set_all_memory(ideal.merged_memory());
4466 set_i_o(ideal.i_o());
4467 set_control(ideal.ctrl());
4468 }
4469
4470 void GraphKit::final_sync(IdealKit& ideal) {
4471 // Final sync IdealKit and graphKit.
4472 sync_kit(ideal);
4473 }
4474
4475 Node* GraphKit::load_String_length(Node* str, bool set_ctrl) {
4476 Node* len = load_array_length(load_String_value(str, set_ctrl));
4477 Node* coder = load_String_coder(str, set_ctrl);
4478 // Divide length by 2 if coder is UTF16
4479 return _gvn.transform(new RShiftINode(len, coder));
4480 }
4481
4482 Node* GraphKit::load_String_value(Node* str, bool set_ctrl) {
4483 int value_offset = java_lang_String::value_offset();
4484 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4485 false, NULL, Type::Offset(0));
4486 const TypePtr* value_field_type = string_type->add_offset(value_offset);
4487 const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::NotNull,
4488 TypeAry::make(TypeInt::BYTE, TypeInt::POS, false, true, true),
4489 ciTypeArrayKlass::make(T_BYTE), true, Type::Offset(0));
4490 Node* p = basic_plus_adr(str, str, value_offset);
4491 Node* load = access_load_at(str, p, value_field_type, value_type, T_OBJECT,
4492 IN_HEAP | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0) | MO_UNORDERED);
4493 return load;
4494 }
4495
4496 Node* GraphKit::load_String_coder(Node* str, bool set_ctrl) {
4497 if (!CompactStrings) {
4498 return intcon(java_lang_String::CODER_UTF16);
4499 }
4500 int coder_offset = java_lang_String::coder_offset();
4501 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4502 false, NULL, Type::Offset(0));
4503 const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
4504
4505 Node* p = basic_plus_adr(str, str, coder_offset);
4506 Node* load = access_load_at(str, p, coder_field_type, TypeInt::BYTE, T_BYTE,
4507 IN_HEAP | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0) | MO_UNORDERED);
4508 return load;
4509 }
4510
4511 void GraphKit::store_String_value(Node* str, Node* value) {
4512 int value_offset = java_lang_String::value_offset();
4513 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4514 false, NULL, Type::Offset(0));
4515 const TypePtr* value_field_type = string_type->add_offset(value_offset);
4516
4517 access_store_at(str, basic_plus_adr(str, value_offset), value_field_type,
4518 value, TypeAryPtr::BYTES, T_OBJECT, IN_HEAP | MO_UNORDERED);
4519 }
4520
4521 void GraphKit::store_String_coder(Node* str, Node* value) {
4522 int coder_offset = java_lang_String::coder_offset();
4523 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4524 false, NULL, Type::Offset(0));
4525 const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
4526
4527 access_store_at(str, basic_plus_adr(str, coder_offset), coder_field_type,
4528 value, TypeInt::BYTE, T_BYTE, IN_HEAP | MO_UNORDERED);
4529 }
4530
4531 // Capture src and dst memory state with a MergeMemNode
4532 Node* GraphKit::capture_memory(const TypePtr* src_type, const TypePtr* dst_type) {
4533 if (src_type == dst_type) {
4534 // Types are equal, we don't need a MergeMemNode
4535 return memory(src_type);
4536 }
4537 MergeMemNode* merge = MergeMemNode::make(map()->memory());
4538 record_for_igvn(merge); // fold it up later, if possible
4539 int src_idx = C->get_alias_index(src_type);
4540 int dst_idx = C->get_alias_index(dst_type);
4541 merge->set_memory_at(src_idx, memory(src_idx));
4542 merge->set_memory_at(dst_idx, memory(dst_idx));
4543 return merge;
4544 }
4617 i_char->init_req(2, AddI(i_char, intcon(2)));
4618
4619 set_control(IfFalse(iff));
4620 set_memory(st, TypeAryPtr::BYTES);
4621 }
4622
4623 Node* GraphKit::make_constant_from_field(ciField* field, Node* obj) {
4624 if (!field->is_constant()) {
4625 return NULL; // Field not marked as constant.
4626 }
4627 ciInstance* holder = NULL;
4628 if (!field->is_static()) {
4629 ciObject* const_oop = obj->bottom_type()->is_oopptr()->const_oop();
4630 if (const_oop != NULL && const_oop->is_instance()) {
4631 holder = const_oop->as_instance();
4632 }
4633 }
4634 const Type* con_type = Type::make_constant_from_field(field, holder, field->layout_type(),
4635 /*is_unsigned_load=*/false);
4636 if (con_type != NULL) {
4637 Node* con = makecon(con_type);
4638 if (field->type()->is_inlinetype()) {
4639 con = InlineTypeNode::make_from_oop(this, con, field->type()->as_inline_klass(), field->is_null_free());
4640 } else if (con_type->is_inlinetypeptr()) {
4641 con = InlineTypeNode::make_from_oop(this, con, con_type->inline_klass(), field->is_null_free());
4642 }
4643 return con;
4644 }
4645 return NULL;
4646 }
4647
4648 //---------------------------load_mirror_from_klass----------------------------
4649 // Given a klass oop, load its java mirror (a java.lang.Class oop).
4650 Node* GraphKit::load_mirror_from_klass(Node* klass) {
4651 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset()));
4652 Node* load = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
4653 // mirror = ((OopHandle)mirror)->resolve();
4654 return access_load(load, TypeInstPtr::MIRROR, T_OBJECT, IN_NATIVE);
4655 }
|